Autologous chondrocyte implantation in the knee: systematic review and economic evaluation

Lavage and debridement

In lavage, the arthroscope (a sort of fibreoptic telescope) is inserted into the knee and saline is poured in through a cannula. This is usually done under general anaesthesia on a day-case basis. The saline washes out loose debris through the cannula. It is also thought to wash out compounds that cause inflammation.

Debridement is done under arthroscopic vision, and is the removal of damaged cartilage or bone.

Debridement and lavage are often done together.

The evidence for effectiveness is sparse and mixed. One three-armed RCT – of lavage alone, lavage plus debridement and a sham arm – reported no difference at 2 years. 18 Another by Hubbard19 had methodological weaknesses, but reported that debridement and lavage was better than lavage alone. The NICE intervention procedures guidance (IPG230)20 noted uncertainty about the efficacy of the procedure.

Autologous chondrocyte implantation

Cartilage cells are called chondrocytes. In ACI, a small piece of cartilage is removed from the knee, and the chondrocytes are grown in the laboratory until they number millions. They are then put onto the damaged area of articular cartilage as a patch. The hope is that this patch will repair the damaged area and form a new layer of natural articular cartilage, called hyaline cartilage.

Autologous chondrocyte implantation has been used since at least 1987,21 and the procedure has evolved over time. The Dutch Orthopaedic Association has provided a useful summary of developments. 22 In the first generation of ACI, the cultured chondrocytes were placed in the defect, in liquid form, and then covered with a cap made from periosteum [autologous chondrocyte implantation–periosteal flap (ACI-P)]. This led to problems with pain in the immediate postoperative period and a need for further procedures to remove overgrowth in the graft as described in Box 1.

The second generation of ACI used a collagen cap [autologous chondrocyte implantation–collagen cap (ACI-C)] instead of the periosteal one, but still used cells in a liquid. Gomoll et al. 24 compared two cohorts, one that had a periosteal patch (ACI-P) and one that had a collagen cap (ACI-C). The reoperation rates were 26% and 5%, respectively. ACI-P is now little used in the UK, but is still used in the USA, where none of the membranes or scaffolds used in second-generation ACI has yet been approved by the Food and Drug Administration (FDA) except for in trials. 25

In the third generation of ACI, the chondrocyte cells are loaded or embedded, or ‘seeded’, on to a porcine collagen membrane ACT-C (autologous chondrocyte transplantation seeded collagen membrane) or matrix [matrix-applied chondrocyte implantation (MACI^®)], with a patch cut to fit. These patches can be implanted by a less-invasive form of surgery, by arthroscopy or mini-arthrotomy, requiring less surgical time than ACI-C. 26 (Arthrotomy = opening of a joint.) ChondroCelect cells (TiGenix, Leuven, Belgium) are now used in this way, with cells being loaded into the membrane by the surgeon.

The membrane used in MACI is composed of type I/III collagen, with a rough side wherein the chondrocytes are seeded and a smooth side that faces into the joint cavity. 26 The membrane is tough enough to be cut to shape or stitched in place, though it is more often glued in place. 26 The membrane is biodegradable. The term ‘scaffold’ is often used instead of membrane. However, the membrane needs careful handling to minimise chondrocyte death during implantation. 27

Another development, which can apply to both second- and third-generation ACI, has been that only selected chondrocytes are used – this is called characterised chondrocyte implantation (CCI). Cells that are most likely to produce hyaline cartilage with predominantly type II collagen, rather than a less resilient cartilage called fibrocartilage, which produces mainly type I collagen,28 are identified during CCI using a panel of biomarkers, including collagen. TiGenix used six biomarkers and Genzyme (Sanofi) also used additional assays in CCI. 29

Table 1 summarises the generations of ACI. (Note: different authors use ‘second generation’ in different ways.) It is worth noting that graft hypertrophy can occur with second- and third-generation ACI. Niethammer et al. 30 reported graft hypertrophy on magnetic resonance imaging (MRI) in 11 of 44 patients who had MACI (with Novocart – TETEC AG, B. Braun, Reutlingen, Germany).

Harris et al. 31 carried out a systematic review of failures and complications after ACI and reported that failure rates were higher with first-generation ACI-P than with second-generation ACI-C.

The Medical Services Advisory Committee (MSAC) report concluded that the ideal application of ACI would be in a full-thickness chondral defect surrounded by healthy cartilage in an otherwise healthy knee. 32

Microfracture

The main alternative method of repair is called ‘microfracture’, in which small holes are drilled through the surface of the bone in the area of damaged cartilage. This allows bleeding from the bone marrow, and the blood carries stem cells into the area where the damaged cartilage has been debrided. These cells form scar cartilage called fibrocartilage, composed of type I collagen. This is regarded as being inferior to hyaline cartilage, being less hardwearing and not expected to last as long. 33

MF may be combined with the insertion of a collagen membrane to cover the MF clot, known as augmented MF.

MF can be done arthroscopically (i.e. without opening the knee joint) and could be done at the same time as washing out a knee joint and stabilising loose tissue (debridement and lavage).

A search of the NICE website found no guidance on MF.

Mosaicplasty

Another method, which is now much less common, is mosaicplasty, sometimes called OATS (osteochondral autograft transfer system), which involves transplanting small sections of cartilage and underlying bone from a less-weight-bearing part of the knee into the damaged area. The pieces are in little cylinder shapes and, once transplanted, have an appearance not unlike a mosaic – hence the name. Mosaicplasty can be used only for small areas of damage (less than 4 cm²) because the transplanted sections have to come from elsewhere in the knee, usually the trochlea. (In some countries, allograft cadaver donor tissue is used, but this does not appear to happen in the UK.)

Mosaicplasty was reviewed by NICE through the Interventional Procedures Programme. 34 The guidance is reproduced in Box 2. It was dated ‘March 2006’ and so may now be out of date.

Mosaicplasty appears to be little used now. In the ACTIVE (Autologous Chondrocyte Transplantation/Implantation Versus Existing Treatment) trial35 (described in Chapter 4) of ACI versus standard methods, such as MF and mosaicplasty, few surgeons chose mosaicplasty.

Conservative management

Another option is no surgical treatment. Three case series36–38 reported high levels of return to activities after cartilage injuries after 14, 9 and 9 years, respectively. Messner and Maletius36 reported a case series of young athletes (mean age 25 years, range 14–38 years) who had no treatment. Fourteen years later, most (21 out of 28) had returned to activity and 22 had excellent or good function. 36 However despite lack of symptoms, most showed radiological changes suggestive of early OA.

The British Association for Surgery of the Knee (BASK) UK Consensus recommends that all patients being considered for ACI should have had physical therapy first, as that may relieve symptoms. 39

Patient group

The patient group, as stated in the final scope from NICE, is ‘adults with a symptomatic cartilage defect (chondral defect) but without advanced osteoarthritis’. Advanced OA is not defined in the scope. The chondral defects can be on the femur, tibia or patella. ACI is used in other joints, but such use is outwith the scope of this appraisal.

No age restriction is given in the scope from NICE, but in past trials, patients had a mean age of 32 years, range 16–49 years, with about 60% men. In most cases, the cartilage damage was due to injury, usually from sport.

Following a UK Cartilage Consensus meeting in March 2014, BASK produced a consensus document. 39 The points most relevant to this appraisal are summarised in Box 4.

Although the consensus recommends increased use of ACI, it would restrict it to people with symptoms and higher-grade lesions. As the statement recognises, some people may have symptoms relieved by physiotherapy. However, physiotherapy cannot repair chondral defects, so this group will still be at risk of progression to OA.

Inclusion criteria

Results

The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram for systematic reviews is shown in Appendix 2.

The quality assessment of the reviews is reported in Appendix 3.

Twelve relevant systematic reviews were included. One by Vasiliadis et al. 45 was associated with a Cochrane review46 but the former provides an update with more trials and is used here. The majority of reviews was rated as ‘at least medium quality’, with three reviews being rated as ‘low quality’,47–49 six reviews rated as ‘medium quality’ (Bekkers 2009,50 Kon 2009,51 Magnussen 2008,52 Mithöfer 2009,53 Nakamura 2009,54 Negrin 201355) and three reviews rated as ‘high quality’ (Harris 2010,44 Vasiliadis 2010,45 Vavken 201056).

Table 52 in Appendix 4 shows the primary intervention studies included in the reviews. Several reviews treated separate publications from the same study, or of subgroups of a study, as separate studies. We therefore checked the original studies and in the table we have grouped all reports from each study together. The tables describing the characteristics of the reviews also record publications from the same study.

The 12 reviews included 27 papers from 19 studies. Eleven of the studies were randomised trials (RCTs) and eight were comparative cohort studies or non-randomised/quasi-randomised trials. None of the primary studies was included in all of the reviews. Of the included primary studies, one compared collagen-based ACI with periosteum-based ACI, four compared ACI with MACI, one compared open with arthroscopic ACI, three compared ACI with mosaicplasty, eight compared ACI with MF, and one each with bone marrow-derived mesenchymal stem cell therapy and with abrasion.

Methods

Basad et al. 2010

This RCT compared MACI, a third-generation ACI (then a Genzyme product) against MF in patients with symptomatic cartilage defects. Patients in the trial came from one centre (the principal author’s clinic in Germany) between 2000 and 2005.

Saris et al. 2014

This trial100 (SUMMIT, Superiority of Matrix-induced autologous chondrocyte implant versus Microfracture for Treatment of symptomatic articular cartilage defects) was a prospective, open-label, parallel-group, multicentre (16 European sites) RCT comparing Genzyme MACI (Genzyme, Europe) against MF.

Search strategy

Searches, as shown in Appendix 1, were run in Ovid MEDLINE and Ovid EMBASE from 1997 to 15 May 2015. Thereafter, weekly auto-alerts, in MEDLINE and EMBASE, of these searches, were run until the end of 2015 to check for any new potential inclusions.

The searches retrieved 2907 documents; after removing duplicates and animal studies, 1833 records remained and the title and abstracts were screened by two authors for inclusions. The full text of 69 articles was checked: 26 articles (21 studies) were included and 43 articles were excluded.

Quality assessment used the National Institutes of Health (NIH) checklist for observational studies. 105

Caveats

When considering survival curves extrapolated beyond the observed data, it should be borne in mind that the extrapolation assumes that the curve based on the observed data will continue. When using parametric fits for extrapolation (any fit irrespective of the equation that describes it), the usual option is to select what is considered to be the best fit to the observed data. Unfortunately, there is no universally applicable method to determine the best fit and opinions may differ. With some data, most well-fitting models will produce similar extrapolations; however, with ACI data this has not been the case. Selecting several plausible models that produce different extrapolations should bracket what can be argued to be the best estimate of behaviour beyond the observed data.

In most instances it was possible to reconstruct individual patient data (IPD) using published data; in the case of degenerative change reported by Nawaz 2014,80 the only way to get parametric curve parameters was using a digitised KM (rather than IPD) with the non-linear regression Stata^® (StataCorp LP, College Station, TX, USA) command specifying candidate parameters. Since in this instance it was not possible to get a meaningful covariance matrix these particular results could only be used in deterministic cost-effectiveness analysis and not for probabilistic sensitivity analyses (PSAs). Occasionally with some other parametric models no confidence interval (CI) could be derived (probably due to lack of numbers for patients and for events) and so uncertainty was not quantifiable (see Appendix 9); again in these instances for cost effectiveness estimates only deterministic analysis was possible.

Methods

Published KM graphs were digitised and IPD reconstructed using the Guyot et al. method. 131 Where published graphs had the appearance of KM plots but authors did not specify their method, it was assumed to be KM. Where plots were presented as scatter graphs rather than stepped lines it was assumed data points represented the top, rather than bottom or midpoint, of a stepped fall in survival.

Parametric models of time to failure were used to explore failure rate beyond the observed data. In the absence of patient numbers this was done by least squares non-linear regression. Where patient numbers allowed use of the Guyot method to reconstruct IPD, the models were implemented in Stata (version 12) with the ‘streg’ command and or using the ‘stgenreg’ package of Crowther and Lambert 2013. 132 Standard models (exponential, Weibull, log-logistic, log-normal, Gompertz and gamma) were explored together with additional models for increasing hazard through time, either after an initial phase of decreasing hazard (bathtub model) or with linear increase in hazard (Rayleigh models). CIs (95%) were estimated with the delta method. The bathtub model was investigated because it has previously been found useful for modelling failure rates after total hip replacement. 133

Linearly increasing hazard models were tried because MF failure rate during 12 years of follow-up of patients with OA (Bae et al. ,134 an excluded study for this report, because patients had OA and were older) was found to be best fit with such models. They were therefore judged to be worth investigating. Linear hazard models were used with one or two parameters in which survival is described as:

(lambda0 ≥ zero; when lambda0 is zero, S conforms to equation 1.)

Model fit was judged using information criteria and by visual inspection of cumulative hazard plots135 and KM plots. One study provided Cox multivariate regression analysis of patient subgroups. The hazard ratios (HRs) from these analyses were used to estimate subgroup failure rates using two methods: (1) A log-normal model hazard was calculated for the baseline subgroup and multiplied by the appropriate HR for each of the other subgroups. The time to failure for these subgroups was estimated from exp(–cumulative hazard). (2) A Weibull or linearly increasing hazard model survival for the subgroups was estimated from: exp(ln (baseline subgroup survival) × HR).

Log-normal model hazard was calculated from:

where:

where Ф is the standard normal distribution.

Parametric models and KM plots are presented in Appendix 9. The remit for this report was to exclude studies with less than 5 years of time-to-event data; therefore, except in exceptional circumstances, such studies were excluded.

Description of time-to-failure data

Seven relevant published studies43,67,80,111,115,117,136 were identified that presented KM plots extending to at least 5 years. Estimates of IPD reconstructed from such plots are best served (i.e. likely to be more accurate) when the total number of events, and of patients, are reported; when accompanied by a risk table indicating the number of participants remaining at risk at multiple time intervals; and when the graphical display is of sufficient quality to unambiguously identify the times at which events occurred. Extrapolation of parametric models’ fit to such IPD are more likely to be reliable for more mature data, that is when follow-up is sufficient that the number of events has reduced the probability of survival at the end of the plot to a low value. A rough estimate of maturity is whether or not median survival has been reached.

Some of the included studies80,136 aggregated event times to yearly intervals. This may be because precise times of failure were not recorded. The impact of this on subsequent use of data is difficult to gauge. Risk tables were rarely presented, and in some studies reporting subgroup analyses the number of patients, as well as the number of events, for some subgroups was not reported. Median survival was reached in only one study. One study136 presented KM analyses in scatter plots rather than a more conventional stepped plot. In this study, it was difficult to be certain whether data points represented the top, bottom or midpoint of a step in survival. These characteristics are summarised in Table 6.

The definition of treatment failure varied between studies; failure either consisted exclusively of surgical re-intervention or a mixture of surgical re-intervention and a poor functional or pain score relative to pre treatment. Table 7 summarises the treatment failure definitions used in the relevant studies.

Results of autologous chondrocyte implantation trials

Four small RCTs, each with two arms were identified. Two (Knutsen 2007,67 N = 80 patients; Vanlauwe 2011,43 N = 112 patients) compared ACI with MF. The RCT of Bentley et al. 201278 (N = 100 with about 10 years of follow-up) compared ACI with mosaicplasty – we are interested only in the ACI arm and these patients were included in the larger study of Nawaz et al. 80 and so are not considered further. The RCT of Gudas et al. 128 compared MF with mosaicplasty but did not satisfy inclusion criteria.

Figure 1 summarises the reconstructed time-to-failure KM plots for the Knutsen et al. 67 and Vanlauwe et al. 43 RCTs, which extended to 5 and 6 years of follow-up, respectively.

FIGURE 1.

Reconstructed KM plots (95% CI) of time to failure in two RCTs.

In these small studies the observed data are associated with considerable uncertainty. Extrapolation of parametric models was associated with large uncertainty beyond the observed data. Figure 2 illustrates this for Weibull fits extrapolated to 50 years (other models are presented in Appendix 9). The short follow-up and small size limits their usefulness for modelling failure rates beyond 5 years. Additional (non-RCT) studies of larger size and longer follow-up were sought.

FIGURE 2.

Extrapolated Weibull distributions fit to reconstructed IPD from the RCT studies. Note: there is some doubt about the reliability of the Vanlauwe et al. 43 MF arm because of an anomaly in the published risk table, which coincided with a flattening of the KM plot at 3 years. Saris et al. 200969 reported 3-year MF data for this trial.

Results of autologous chondrocyte implantation observational studies

Four single-arm ACI studies with KM plots were included. Niemeyer et al. 117 reported event times for 70 German patients with follow-up to 5 years; Minas et al. 136 and Moseley et al. 115 reported time to failure for 210 and 72 US patients, respectively, with follow-up extending up to or beyond 10 years; and Nawaz et al. 85 reported annualised time to failure for 827 UK patients, with follow-up to about 10 years. ACI patients from Bentley et al. 78 and from Biant et al. 79 were included in the study of Nawaz et al. 80 and so survival data from these are not considered separately. For the whole cohort, Nawaz et al. 80 reported both annual event and censoring numbers for each year so that the IPD could be reconstructed without resort to the Guyot et al. 131 algorithm. Jungmann et al. 111 presented time to re-intervention for 413 of 500 patients (selected follow-up 2–11.8 years) with analysis truncated at 5 years; this KM was not comparable with those in other studies and IPD were not reconstructed.

Patient characteristics in the ACI studies are summarised in Table 8. Typically, lesions were full thickness, with study mean size ranging from 2.7 to 8.4 cm² in patients with mean age 30–40 years, most of whom had experienced previous interventions. Symptom duration prior to intervention varied between studies.

Figure 3a shows the reconstructed KM failure plots for the four single-arm and two RCT ACI studies that provide relevant data to at least 5 years. It should be appreciated that definitions of failure were not identical between different studies, and the mix of patients who had, or had not, experienced previous intervention also differed between studies. Because of study size, the uncertainty in the Nawaz et al. 80 data is less than that in the other studies. Up to about 6 years there is reasonable consonance for most studies; thereafter, the prognosis appears worse for Nawaz et al. 80 patients, but this appears partly due to flat portions of the KM curves for the other studies for which patients at risk have diminished considerably and uncertainty is at its maximum. These data indicate that the Nawaz et al. study80 is unlikely to flatter failure rates after ACI, and that to about 6 years of follow-up the Nawaz et al. study80 is reasonably consistent with other studies; beyond 6 years the KM analyses of the other studies are likely to be less reliable because of smaller study size. Combination of these different studies would be difficult to justify because of clinical heterogeneity.

FIGURE 3.

(a) Reconstructed KM plots for ACI studies; and (b) best parametric fit (95% CIs) and gamma models for six ACI arms.

Figure 3b shows parametric models for these studies extrapolated to 70 years. For clarity Moseley115 is omitted. A gamma model could not be computed for Minas et al. 136 Gamma fits illustrate differences seen between studies when applying the same single distribution to all; also shown are best fits for each study. Judged according to information criteria, various parametric models provided best fits: Knutsen67 exponential, Minas136 Gompertz, Moseley115 exponential, Nawaz80 log-logistic, Niemeyer117 exponential and Vanlauwe43 linear hazard. For Nawaz80 and Vanlauwe43 the best fits differed very little from the gamma model. The best fit models for Knutsen, Nawaz,80 Niemeyer117 and Vanlauwe43 studies predict that more than half of ACI interventions fail within about 30 years, whereas the Gompertz model-based Minas study136 indicates about half or more ACIs remain without failure up to 70 years, and the gamma model for the Moseley study115 predicts about 25% remain without failure to 70 years.

Potential reasons for differences in KM plots and best fit models between these studies are manifold; they include uncertainty in the observed data resulting from small numbers of participants and, in some studies, short-term follow-up, different reliability of IPD reconstructions and differences in study populations, particularly with regard to experience of previous intervention(s), the degree of degenerative change, and location and size of lesion.

Nawaz et al. 2014 study of UK patients

The most useful study is by Nawaz et al. 80 For the whole Nawaz cohort (n = 827)80 a log-logistic distribution provided the best-fitting parametric model. Figure 4 shows the reconstructed KM plot together with the log-logistic model extrapolated to 50 years; the model predicts that after about 30 years approximately 90% of patients would have failed. The partition of failures according to how it was defined (see Table 7) was not reported (e.g. the proportion of failures receiving previous intervention at the time of failure is unknown).

FIGURE 4.

Reconstructed KM plot (a) and extrapolated log-logistic model; and (b) for the Nawaz et al. 80 whole cohort.

Patient subgroups examined in the Nawaz study

Nawaz et al. 80 presented KM plots for subgroups of patients categorised according to (1) receipt of a previous intervention; (2) site of intervention; (3) grade of degenerative change; and (4) type of intervention received (MACI or ACI). The authors used univariate and multivariate Cox regression to investigate if these – and also age and size of defect – were influential for failure. The most influential patient covariate was previous intervention (p < 0.001; multivariate HR vs. no previous intervention: 4.72, 95% CI 3.5 – 6.4). Grade of degenerative change (p < 0.001), site of intervention (p = 0.036 for best vs. worst site), and age at operation (p < 0.001) were also significantly influential, whereas type of intervention (ACI or MACI) and lesion size were not (p = 0.860 and p = 1.00, respectively). The authors did not report on a test of the proportional hazards assumption. The AG reconstructed the subgroup KM plots and used reconstructed IPD to investigate good parametric models for the data. Additionally, AG investigated the effect of adjusting parametric models using the multivariate HRs reported by Nawaz et al. 80

Previous and no previous intervention

According to information criteria, log-normal and gamma distributions provided good models for patients who had previous or no previous intervention (debridement was not included as a previous intervention). When the HR reported by Nawaz et al. 80 (previous vs. no previous intervention) was applied to either log-normal or Weibull models, the resulting model was very similar to that fit to the previous subgroup IPD (Figure 5). These results indicate that there was likely to be little difference between the subgroups in the distribution of other covariates that were influential for failure.

FIGURE 5.

Reconstructed plots for Nawaz et al. 80 Previous and no previous intervention subgroups: (a) log-normal models; and (b) Weibull models. KM plots and extrapolated log-normal (left) and Weibull (right) parametric models. Dashed lines are 95% CIs.

Site of intervention

Nawaz et al. 80 published KM plots and multivariate Cox regression HRs comparing time to failure for five subgroups that differed according to intervention site (medial femoral, n = 421; lateral femoral, n = 109: patella, n = 200; trochlea, n = 50; multiple sites, n = 47). HRs vs. the lateral femoral condyle group as baseline reference were medial femoral condyle 1.806 (95% CI 1.036 to 3.149; p = 0.037), patella 1.323 (95% CI 0.745 to 2.351; p = 0.339), trochlea 1.409 (95% CI 0.625 to 3.174; p = 0.0408) and multisite 1.678 (95% CI 0.731 to 3.851; p = 0.222). Reconstructed KM plots were similar for all but the lateral condyle group, which exhibited the least failure. Log-normal distributions provided the best fit parametric models to reconstructed subgroup IPD. Figure 6 shows reconstructed KM pots and HR-adjusted log-normal models. Applying the reported HRs diminished the apparent superiority of the lateral femoral condyle subgroup seen in the KM plots, and indicated that relative to other subgroups the lateral femoral population may possibly have been favourably free of detrimental covariates for failure (e.g. previous treatment and high-grade degenerative change). Similar results were obtained with Weibull models.

FIGURE 6.

Reconstructed KM plots and log-normal models for time to failure according to site of intervention, Nawaz et al. 80 (a) KM according to intervention site; and (b) subgroups by intervention site log-normal models based on HRs. For clarity, KM 95% CIs are shown only for lateral and medial femoral condyle sites and have been omitted for the model fits.

Grade of preoperative degenerative change

Nawaz et al. 80 published KM plots and multivariate Cox regression HRs comparing time to failure for four subgroups categorised according to grade of degenerative change (Kellgren–Lawrence). HRs vs. the grade 0 subgroup as baseline reference were grade 1, 1.542 (95% CI 0.930 to 2.557; p = 0.093); grade 2, 1.869 (95% CI 1.381 to 2.529; p < 0.001); and grade 3, 1.985 (95% CI 1.092 to 3.610; p = 0.025). Patient numbers were not reported and parametric models were fit to digitised KM plots using non-linear regression. For different subgroup grades, log-normal and linearly increasing hazard models produced acceptable fit to digitised KM plots in Appendix 9. When the HRs reported by Nawaz et al. 80 were applied to either of these models, the resulting plots for different grades were more similar to each other than was apparent from KM plots or fits to KM plots (Figure 7). These results may indicate that some of the superiority of the grade 0 subgroup apparent in the KM plots was possibly due to relative freedom of this group from covariates that tend to increase the probability of failure.

FIGURE 7.

Time to failure by grade of OA. (a) Nawaz et al. 80 subgroups by grade of degenerative change linearly increasing hazard models; and (b) Nawaz et al. 80 subgroups by grade of degenerative change log-normal models.

Minas et al. 2014 study

Minas et al. 136 performed several KM analyses for various subgroups of patients. Patient numbers were reported only for the comparison of previous intervention versus no previous intervention groups. Like in the Nawaz et al. study,80 worse failure rates were found for patients who had experienced previous intervention. As was seen for the whole Minas cohort,136 the subgroup failure rates flattened after about 6 years and extended to as far as 17 years with relatively few failure events. Thus failures were much less frequent in both Minas136 subgroups than in the corresponding Nawaz80 subgroups. No regression analysis was performed in Minas136 and no HRs were reported. Gamma distributions provided good fits for both studies’ subgroups. Reconstructed KM plots and gamma models fit to IPD for subgroups from both the Nawaz80 and Minas136 studies are shown in Figure 8.

FIGURE 8.

Effect of previous interventions on failure rates. (a) Nawaz et al. 80 and Minas et al. ,136 no previous intervention populations. Reconstructed KM plots and gamma models; and (b) Nawaz et al. 80 and Minas et al. ,136 previous intervention populations. Reconstructed KM plots and gamma models. Time to failure for previous intervention and no previous intervention patient subgroups (Minas136 and Nawaz80) showing reconstructed KM plots and gamma models of time to failure.

Minas et al. 136 also provided plots for failure according to subgroups that experienced different types of previous intervention. Patient numbers were as follows: MF, n = 13; abrasion arthroplasty, n = 30; and drilling, n = 46. Failure was more frequent after MF than after other forms of marrow stimulation. Concurrent osteotomy resulted in fewer failures.

Studies of failure after microfracture (Layton 2015, Knutsen 2007, Saris 2009)

Vanlauwe et al. 43 (year 3 results in Saris et al. 70) and Knutsen et al. 67 provided MF failure data to 5 years. A large US study by Layton et al. 114 examined records for 3498 US recipients of MF and reported the percentages of failures for patients followed to 1, 3 and 5 years. All patients were followed up to 3 years. The proportion followed to 5 years was not reported. Layton et al. stated ‘Failure rates (TKR, MF or ACI) increased with increasing years of follow-up: 9% within 1 year, 18% within 3 years, and 32% within 5 years’. In the Knutsen,67 Saris70 and Layton114 studies, failure was defined as re-intervention. Only Layton et al. 114 provided information on the type of re-intervention received, as follows: TKR accounted for most re-interventions – 56%, 62% and 66% of re-interventions at years 1, 3 and 5, respectively; MF and ACI accounted for nearly all of the remaining re-interventions (very few re-interventions were OATS). The mean age of patients in the Layton study114 was 47 years [standard deviation (SD) 11.4 years], meaning that many would be of an age at which TKR would be considered, and there were equal numbers by gender. Table 9 summarises the main characteristics of patients in the MF arms of the Knutsen67 and Saris70 studies.

As all of the patients in the Layton study114 were followed up for 3 years, it was possible to reconstruct IPD for an annualised KM plot (assuming failure took place at 1 and 3 years, and at 3 years all non-failed patients were censored). Under the assumption that those followed up for 5 years were representative of all of those who could have been followed (censoring those without failure at 5 years), the 5-year IPD were also estimated. The best fit models (Figure 9) for these were provided by Gompertz distributions and the second best by a gamma model (see Appendix 9).

FIGURE 9.

Best fit models to observed MF failure in three studies.

The linearly increasing hazard model provided the best fit for the MF arm of the Knutsen et al. 67 study.

The published 5-year KM plot for the MF arm of the Vanlauwe et al. study43 had anomalous risk table data and interpretation of the KM plot was problematic. The KM plot for MF has 10 steps and a total of 10 events were reported (one step for each event). Seven steps occur before 36 months: two of these very close together, at about 20 months, and three after 36 months. This does not tally with the data in their appendix 1, which depict five MF re-interventions occurring before 36 months and five after 36 months. For the ACI KM plot, two steps occur before 36 months and five steps after 36 months, and this corresponds to the data provided in appendix 1 of the Vanlauwe et al. study. 43 The risk table for the MF arm is anomalous in that the number at risk is reported as increasing at 36 months. It is unclear what the correct numbers should be at 24, 36 and 48 months for the MF risk table. Therefore, the Saris et al. 70 3-year KM plot for this study was examined. The best fit was again provided by the linearly increasing hazard model. These and models for the Layton et al. study114 are summarised in Figure 9. The poorer performance in the Layton et al. study114 may be attributable to older mean age and or real-world performance of MF relative to that for patients who are carefully selected for a RCT.

Comparison of failure after autologous chondrocyte implantation and microfracture

A comparison of long-term failure of MF and ACI is problematical in view of the paucity and heterogeneity of studies. The most reliable comparison may be between the largest UK extended follow-up study (Nawaz et al. 80) and the available MF data (described above), a caveat being that failure definitions differed between the Nawaz study80 and the three available MF studies. When whole cohorts were compared, ACI appears to be superior to MF (Figure 10).

FIGURE 10.

Modelled failure profiles following ACI or MF. Best fit models for Nawaz et al. 80 whole cohort, (all) log-log; Nawaz et al. ,80 no previous subgroup gamma; Knutsen et al. 67 and Saris et al. ,70 MF arms linearly increasing hazard; and Layton et al. ,114 Gompertz (for clarity not all 95% CIs are shown).

No subgroup data was available from the MF studies. Vanlauwe et al. 43 did not provide KM plots for subgroups but reported the failure numbers according to whether or not previous intervention had been experienced (Table 10). Numbers of patients at risk and numbers of events were small and the time of events in compared groups was not provided so that firm conclusions are impossible; however, these data are suggestive of little effect of previous intervention on risk of failure after either ACI or MF. Salzmann et al. 137 followed 454 recipients of MF and compared patient characteristics between those patients who required re-intervention during follow-up with those who did not require re-intervention. The former patients, on average, had received more pre-MF interventions (1.9 ± 2.1 previous interventions) than the latter (1.2 ± 2.1 previous interventions) but the spread in number of pre-interventions was great in both cases. Unfortunately, no KM time-to-event analyses were reported for the ‘no-previous intervention’ and ‘previous intervention’ subgroups.

In the absence of subgroup KM data for MF, the worst-performing subgroups investigated by Nawaz et al. 80 were compared with the three MF studies. Log-normal models based on the multivariate HRs reported by Nawaz et al. 80 were used for the comparison (Figure 11).

FIGURE 11.

Modelled failure in MF studies compared with worst-performing subgroups from Nawaz et al. 80 ACI study.

Except for the previous intervention ACI subgroup, the ACI subgroups clearly exhibited less failure than the MF cohorts. Lack of data does not allow a comparison with previously treated MF patients. It should be emphasised that uncertainty in these comparisons is substantial, especially with regard to the Knutsen67 and Saris70 MF arms. Analysis of the MF arm of the RCT of Gudas et al. 128 – excluded on the basis of its small size – showed that the best fit for the reconstructed IPD was a log-normal model, which predicted poorer survival than the models for MF for the Layton,114 Knutsen67 and Saris70 studies.

Summary of longer-term time-to-event evidence of treatment failure after autologous chondrocyte implantation and microfracture

• More long-term evidence was available for ACI than for MF.

• Treatment failure definitions differed between studies, with varying, and sometimes unclear, relative contributions to overall failure from re-intervention and from inadequate pain/function scores.

• Study data were generally still too short term. Only one published study allowed an estimate of observed median time to failure.

• Most participants in most study populations had experienced intervention(s) prior to enrolment. When evidence was reported it appears many types of pre-intervention had been tried.

• Two ACI studies with KM analyses extending to at least 10 years reported that treatment failure was far more frequent in patients who had experienced prior intervention(s); one of these studies documented greater failure rates after MF than after other marrow stimulation (but patient numbers were small).

• There was no clear time-to-event evidence that prior intervention influenced failure after MF; other available evidence was meagre.

• Immaturity of failure data necessitated parametric modelling beyond observed data so as to predict lifetime failure.

• According to information criteria and visual goodness of fit, the best fits of long-term failure after ACI were usually characterised by models that, when extrapolated beyond the observed data, indicated gradually decreasing hazard (probability of failure decreasing with time).

• Conversely, good fits to limited data available for MF were characterised by models that indicated linearly increasing hazard (probability of failure increasing with time).

• A single large US study of MF in patients with a mean age of 47 years indicated that, in this population, TKR was the most frequent intervention after failure of MF.

Pooling time-to-failure studies

The second submission from SoBi used parametric models based on pooled data from ACI studies to derive time to failure for ACI. SoBi did not pool MF studies. A commentary on the SoBi submission follows in Chapter 4.

The ACI studies pooled by SoBi encompassed studies using different definitions of failure and recruiting different proportions of previously treated and previously untreated patients. More judicious pooling can be undertaken, in which there is less heterogeneity among pooled studies. Therefore, as a supplement to the analysis of single studies described above, the AG have briefly explored pooling of studies for ACI and for MF.

Microfracture studies

FIGURE 12.

Time to failure after pooling four ACI studies and three MF studies.

In the three studies providing MF data, failure was defined as re-intervention and two predominantly recruited patients who had experienced previous interventions. This was not reported by Layton et al. 114 When the three studies (Layton,114 Vanlauwe43 and Knutsen67) were pooled the resulting KM plot and best fit model (Gompertz) were dominated by the large Layton study114 (see Figure 12). Compared with the pooled ACI studies, failure was more frequent in the MF studies.

The pooled MF studies were dominated by the Layton study. 114 Pooled MF studies, excluding Layton,114 Knutsen67 and Saris,70 again indicated less failure for ACI patients than for MF patients (Figure 13).

FIGURE 13.

Time to failure after pooling four ACI and two MF studies (Knutsen67 and Saris67).

When the MF arms of Knutsen67 and Vanlauwe43 (5-year MF data of the TIG/ACT study43) were pooled, it was difficult to determine the best fit model using information criteria (Table 11). Only the gamma model of MF failure was superior to ACI (Figure 14). It should be noted that anomalies in the published Vanlauwe43 MF arm required speculative interpolation of risk table data prior to pooling.

FIGURE 14.

Time to failure after pooling four ACI and two MF studies (Knutsen67 and Vanlauwe43).

TABLE 11 - Information criteria for models to pooled MF data from Knutsen67 and Vanlauwe43

df, degrees of freedom; Obs., observations.

Model	Obs.	ll(model)	df	AIC	BIC	AIC rank	BIC rank
Gamma	101	–58.4314	3	122.8628	130.7081	1	5
Exponential	101	–61.9545	1	125.9089	128.524	5	1
Weibull	101	–60.7536	2	125.5071	130.7374	4	6
Gompertz	101	–61.6625	2	127.325	132.5553	8	8
Log-normal	101	–59.7245	2	123.449	128.6793	2	2
Log-logistic	101	–60.5069	2	125.0139	130.2441	3	4
Linear hazard, one parameter	101	–62.1378	1	126.2756	128.8907	6	3
Bathtub	101	–61.4594	3	128.9189	136.7642	9	9
Linear hazard, two parameter	101	–61.4594	2	126.9189	132.1491	7	7

Study design, intervention and patient populations

The remaining four peer-reviewed journal articles are summarised in Table 12. One study was a cross-sectional retrospective study (Derrett 2005140) and the other three studies were decision-analytical modelling studies by Gerlier et al. ,141 Samuelson et al. 142 and Koerber et al. 143 The Derrett study140 was conducted in the UK, and the other three studies141–143 were based on literature and some trial data from Belgium, Germany and the USA. Three studies assessed the cost-effectiveness of ACI compared with other interventions: mosaicplasty (Derrett 2005140); MF (Gerlier 2010141); mosaicplasty and MF with different versions of ACI (ACI-C, ACI-P and MACI) (Koerber 2013143). Samuelson et al. 142 compared ACI-C with ACI-P to see whether or not it was more cost-effective.

The patient populations varied. The retrospective study by Derrett et al. 140 was based on 95 patients, of whom 53 patients received ACI, 20 patients received mosaicplasty and 22 patients were on the waiting list for ACI. The patients who received ACI were slightly younger than those who had received mosaicplasty (31.9 years vs. 34.9 years; p = 0.17) and more men received ACI (53% men vs. 47% women) compared with mosaicplasty (45% men vs. 55% women). The three economic models were based on clinical data and data from the literature.

Gerlier et al. 141 compared adult patients who were less than 50 years of age (a mean age of 35 years at model entry), with symptomatic cartilage lesions of the femoral condyles, who had not yet developed OA, and the key efficacy data came from the TIG/ACT trial. 43

Samuelson et al. 142 compared adult patients with a mean age of 30 years with a focal chondral injury which satisfied the conditions for an ACI repair.

The model by Koerber et al. 143 was said to be based on the model by Gerlier et al. 141 In the Koerber supplementary file it was stated that the study population was patients aged 32 years with symptomatic, isolated cartilage defects and no contraindication. None of the economic models specified the number of hypothetical patients used for the modelling.

Time horizon and length of follow-up

The time horizon for any study should be long enough to capture all of the benefits that would accrue from the different interventions. The follow-up length in the studies varied. The Derrett study140 was based on follow-up data for 2 years. The economic model by Gerlier et al. 141 used two time horizons: a short-term time horizon of 5 years to take into account knee pain and mobility after the initial intervention (this information was obtained from a 5-year RCT that compared ACI with ChondroCelect and MF) and a long-term time horizon of 40 years to take into account the development of OA after 15 years and the need for a TKR after 20 years. Samuelson et al. 142 based their model on a 10-year time horizon that corresponded with the longest-term evidence that was available in the literature. Koerber et al. 143 stated that on the basis of the German life expectancy of the patients in the model the time frame was set to 47 years. Although the authors did not explicitly state the cycle length – from the information provided this can be deduced as 1 year. Neither Gerlier et al. 141 nor Samuelson et al. 142 reported the cycle length that was used in the model, and none of the three studies applied a half-cycle correction to the economic models.

Study perspective and outcomes

Study perspective is crucial to the economic evaluation, as it will determine whether the appropriate resource use and costs have been collected, calculated and reported. Only two studies explicitly stated the viewpoint for the economic analysis: Gerlier et al. 141 conducted the study from the perspective of the global health-care payer, whereas Koerber et al. 143 conducted their study from the viewpoint of the German statutory health insurance. All four studies conducted a cost–utility analysis for which the final outcomes were reported as QALYs. In addition, the Derrett study140 used a range of outcome measures to compare the groups after surgery. The postoperative group consisted of patients who received either ACI or mosaicplasty who were compared with the ACI waiting list group. Outcome measures included:

• The Cincinnati knee rating scale, which assesses 11 components, including subjective symptoms such as pain and swelling, and functional activity level such as walking and climbing stairs scores – these scores were higher in the combined surgery group than the waiting list group.

• The Pain Disability Index, which helps patients measure the degree to which their daily lives are disrupted by pain – the authors found that patients in the combined surgery group had less pain than the waiting list group (p = 0.09).

• The generic health-related QoL – EuroQol-5 Dimensions, three-level version (EQ-5D-3L) – measure. Patients in the combined group had statistically significantly higher EQ-5D scores than the waiting list group (0.61 vs. 0.41; p = 0.03). The EuroQol-5 Dimensions (EQ-5D) measure was used to calculate the QALYs.

The Gerlier study141 used data from the SF-36 measure to calculate QALYs (this information was collected over a period of 60 months after randomisation from a RCT); in addition, they also used the KOOS, which evaluates five key dimensions: pain, symptoms, ADL, sport and recreation function, and knee-related QoL. Samuelson obtained utility values from the literature to calculate QALYs, although they did not specifically state which instrument or what method was used to estimate these utility values that were used in the model. In addition, some studies used in the model had used the Lysholm knee score (this measure contains eight domains, with a higher score indicating a better outcome) to estimate the utility values. Koerber et al. 143 obtained from the literature (no information sources were provided) and were based on the following: utility after treatment pain free (high functionality), utility with low functionality of the knee (medium functionality) and utility before knee prosthesis with strong pain (low functionality).

Resource use, costs and discounting

Derrett et al. 140 provided a comprehensive breakdown of resource use and costs, which were collected for the economic evaluation. These included secondary-care resource use related to each procedure, which was collected from patients’ electronic and medical records from the time point of the first preoperative outpatient appointment to 2 years postoperatively. In addition, they also stated price year for which the costing was undertaken (year 2003–4). The resource use and costs of the surgical procedures, and the follow-up costs after initial interventions that were used in the model, have been comprehensively listed by Gerlier et al. 141 This included information detailing the length of stay for each procedure and follow-up stage and also stated the price year for the economic analysis (year 2008). Both the Samuelson142 and Koerber143 studies provided resource use and cost information, but not as detailed as the Derrett140 and Gerlier141 studies. For example, for the different procedures the components were not individually listed and the price years for which the economic analysis was not explicitly stated – therefore researchers cannot use these unit costs for their own studies or to conduct a cost comparison with their own or with other studies.

All three of the economic models performed discounting using both 3% for costs and outcomes, except for Gerlier et al. ,141 who used 1.5% for outcomes. Derrett et al. 140 did not conduct discounting, stating ‘that costs tended to occur in the first year, making discounting unnecessary . . . the exact timing of postoperative benefit accrual was unknown’. Discounting is important in cost-effectiveness analyses, as it converts future costs into present values, thereby allowing comparisons between costs and benefits that occur at different times. This is especially important for different interventions for which costs usually occur in the current time period, whereas benefits are usually not evident until some point in the future; hence, discounting should have been undertaken by Derrett et al. 140 because the study length was greater than 1 year.

Base-case results and sensitivity analyses

The results and the conclusions offered by each study differed: Derrett et al. 140 found that the average cost was higher for ACI than mosaicplasty (£10,600 vs. £7948 in 2003–4 prices). Outcomes in terms of EQ-5D were better for the ACI group than with mosaicplasty (0.64 vs. 0.47); this difference was not statistically significant. Overall, the ICER for providing ACI relative to mosaicplasty was £16,349.

Gerlier et al. 141 found that the mean costs of ChondroCelect ACI were higher than MF (€29,808 vs. €9006 in 2008 prices), but the overall mean QALYs were also higher for the ACI group (21.08 vs. 19.79). The authors found that the probability of ACI being cost-effective was approximately 80% if the payer has a willingness to pay of €22,000 per QALY. The cost per QALY gained for ACI over MF was €16,229.

Samuelson et al. 142 found that the total costs of ACI-C were slightly higher than ACI-P – a difference of US$188 (US$66,940 vs. US$66,752). However, there was some conflicting evidence when they later say that ACI-C was less expensive by US$941. The earlier figure we presume relates to the initial cost difference and the latter figure must be after the model was run for 10 years; however, this was not explicitly stated. Also, no further information or breakdown was provided by the authors to show how these costs were obtained or calculated. Individual QALY means were not reported over the 10-year period, except the authors stated that ACI-C was more effective by 0.07 QALYs. The authors calculated a cost per QALY for each of the two different ACI interventions by dividing the cost of the intervention by the QALY to get a cost per QALY; however, this was not an incremental cost. Also, we could not work backwards to find out what these individual costs and QALYs were for each intervention. From the information gleaned from the paper, the ICER should have been reported as the cost per QALY gained of ACI-C relative to ACI-P – US$13,443 (US$941/0.07).

Koerber et al. 143 reported mean costs and QALYs for each intervention separately; the costs ranging from €13,445 (MF) to €21,204 (MACI), and QALYs ranging from 19.47 (mosaicplasty) to 19.80 (MACI). The cost per QALY gained was worked out for each intervention in relation to MF; the authors found that mosaicplasty was dominated by MF (MF was cheaper and more effective), whereas the cost per QALY gained ranged from €40,523 for ACI-C to €56,370 for ACI-P both in relation to MF.

Sensitivity analyses are important in economic analyses, as they deal with uncertainty around key parameters and assumptions made in the model, and help confirm the robustness of the results. All four studies conducted some sort of sensitivity analyses (SAs), ranging from the most simplistic one-way SAs (Derrett140) to the more sophisticated probabilistic analyses (Gerlier 2010,141 Koerber 2013143).

Limitations of the studies

All four articles had some methodological shortcomings. For example, in the Derrett study,140 patients were not randomly assigned to treatment groups and follow-up was for only 2 years. The perspective of the economic analysis was not stated, and both costs and benefits were not discounted; only one SA was carried out, which looked at the lowering the costs of the ACI (where the ICER decreased slightly), and there were no preoperative utility scores for both groups (therefore utility values from a waiting list group were used). Gerlier et al. 141 felt that there were not enough data on the probability and time to occurrence for specific events such as TKR, which meant that a Markov model could not be developed. Another key limitation was the lack of long-term clinical follow-up data that could be used in the model, but one of the strengths of the study was the use of the data from the RCT to help populate the model. The limitations in the Samuelson study142 are most notably the inability to calculate the ICER (cost per QALY gained of ACI-C relative to ACI-P) accurately; short follow-up (10 years); perspective of the economic analysis was not stated; lack of trial data and the model relied heavily on assumptions and data from different studies in the literature; and lack of data on the QoL, that is the authors assumed utility values after both ACI-C and ACI-P were the same, as were the failure rates. Koerber et al. 143 did not explicitly evaluate ACI, but merely used ACI as an example to explain early evaluation and value-based pricing of regenerative medical technologies, although they did provide a supplementary file with some of the model inputs.

Quality assessment

The quality of the reporting of the economic analyses by the four articles was assessed using the 27-point CHEERS checklist. 138 Koerber et al. 143 did not identify the study as an economic evaluation in the title, nor did they provide a structured abstract. Only two studies reported the viewpoint of the economic analysis (Gerlier 2010,141 Koerber 2013143). Samuelson et al. 142 did not describe all of the comparators fully. The choice of health outcomes was well reported by all four studies; in terms of analytical methods and study parameters, these were best reported by Derrett et al. 140 and Gerlier et al. 141 The article by Gerlier et al. 141 was the most comprehensively completed in terms of economic analysis using the CHEERS checklist: 18 of the 27 statements (66.7%) were a ‘yes’, four statements (14.8%) were partially completed, two statements (7.4%) were not completed and three statements (11.1%) did not apply. The least comprehensive article in terms of the economic analysis was the article by Koerber et al. ,143 in which their study resulted in ‘yes’ to only 7 out of the 27 statements (25.9%); eight statements were partially completed (29.6%), five statements were not completed (18.5%) and three statements did not apply (11.1%).

Using the adapted Phillips et al. 139 32-point checklist to critically appraise the economic models, overall the four articles adequately reported the objective of the model evaluation, the structure of the model, the type of model for the decision problem, the methods and assumptions to extrapolate short-term results into final outcomes, and the costs used in the model. The models did not provide clear justification if any feasible options were excluded. The cycle length was not explicitly stated in any of the studies, the choice of baseline data was not justified and none of the methods used expert opinion. None of the models applied a half-cycle correction, or justified that omission. Again, the Gerlier article141 was the most comprehensive analysis when using the Philips checklist139 to critique the article: 21 out of the 32 statements (65.7%) were a ‘yes’, five statements (15.6%) were partially completed and six statements (18.8%) were not completed. The Samuelson article142 was not as comprehensively completed in terms of the economic model: only 8 of the 32 statements were a ‘yes’ (25.0%), 10 statements (31.3%) were only partially completed and nine statements were not completed (28.1%).

We also note an Austrian HTA report by Kunzl et al. 145 from the Ludwig Boltzmann Gesellschaft HTA unit, which commented that Austria was one of the few countries that funded ACI. However, the Ludwig Boltzmann Gesellschaft HTA report concluded that in 2009 there was a lack of evidence that ACI was more clinically effective than the other options. No cost-effectiveness analysis was performed.

Trial data: autologous chondrocyte implantation–periosteal flap versus microfracture: TIG/ACT/01/2000

This trial43 compared ACI-P with CCI against MF in patients with symptomatic cartilage defects of the femoral condyles. The 5-year results are reported by Vanlauwe et al. 43 Other papers from this study include Saris 200869 and Saris 2009. 70 The former provides 12- and 18-month follow-up results and the latter has 36-month follow-up results.

Case series

The baseline characteristics of patients in the case series were more varied in some ways than in the RCT, as shown in Table 15.

The outcomes in the compassionate case series were the Clinical Global Impression measures of improvement (CGI-I) and efficacy (CGI-E). CGI-I measures change from baseline (no change, improvement, worsening). CGI-E has four points: very good, moderate, slight and no change or worse. Results were divided into short-term follow-up (under 18 months, mean 9 months, which is too short for best outcome) and longer term (> 18 months, mean 27 months) but figures in each group are not given.

Note that these scales are reported by the surgeon, not the patient. The CGI-I results were reported as showing good outcomes (much improved or very much improved) in 68%, with serious worsening in only 2%. The CGI-E results showed 38% with very good results, 36% with moderate improvement, 12% with slight improvement and 11% unchanged or worse. (From submission table 10, p. 30 – results total 97% not 100%.)

The submission reports that no differences were seen by duration of follow-up (< 18 months vs. > 18 months), site of lesion (patella vs. condyles) or size of lesion (small vs. large, not defined). Patients with single lesions did better than those with multiple ones, but only significantly so in CGI-I results (improved 86% vs. 77%). Results in multiple lesions were good.

The most common AE was knee pain (24%) and 54% had no AEs. As expected with the ACI-C method, few patients (2%) developed cartilage hypertrophy.

Registry cohort

Details from this cohort are sparse and only about half of the cohort (153 out of 308) had 6 months or more of follow-up. The mean age of 32 years (range 15–50 years) is similar to the RCT and case series. The only benefit reported is an increase in KOOS, at up to 36 months, but numbers at each follow-up period are not given. There were six treatment failures, and two DVTs among a total of 17 SAEs (but no denominator given). Treatment failure was defined as the need for a re-intervention for more than 20% of the treated area, associated with symptoms. The summary states that no new AEs were reported in the registry cohort.

Belgian reimbursement scheme

Little information is reported from this source. Two procedures failed within 12 months and another two procedures failed between 12 and 24 months, in 254 patients. Only 51 patients had reached 3 years of follow-up.

The data show an increase in numbers treated, from 51 in the first year (May 2011 to April 2012), 93 in the second and 110 in the third, possibly suggesting levelling off in numbers. The population of Belgium is 11.2 million, so the third year rate is about 10 per million per year. The equivalent numbers per year in England would be 540, and in Wales 30.

The ChondroCelect submission argues, with some justification (see Chapter 2 of this report), that ACI is more successful as a primary procedure in patients who have had previous MF. The Minas et al. study75 is cited in support of the assertion.

Cost-effectiveness

The health-related QoL was measured in the TIG/ACT trial43 using the SF-36 administered at 18, 24, 30, 36, 48 and 60 months post procedure. At 36 months, SF-36 scores were slightly better for ACI.

Cost-effectiveness

The submission by Aastrom Biosciences did not provide any cost-effectiveness analyses because of the recent purchase of the MACI product by Vericel from Sanofi. Cost-effectiveness evidence was presented in the MSAC submission32 and the manufacturers aimed to adapt this. However, this was not possible because of time constraints, so only a budget impact/costing forecast model was provided.

The budget impact model estimated by Aastrom indicated that 9549 patients in England and Wales were eligible for cartilage repair in 2013. Of these 9549 patients, as indicated by the NICE scope, 500 will be eligible for MACI or an ACI in year 5. The manufacturers assumed that there would be an equal split of the use of MACI and ACI. The rest of the eligible patients would receive MF, though the reasons for not offering ACI are not explained. Based on data from three studies (Minas 2009,75 Nawaz 2014,80 Vijayan 2014153) the manufacturers reasonably assumed that reoperations after MF do not have the same success rate as primary MACIs or ACIs.

List prices were used for the costs for ACI (£18,300) and MACI (£16,226 excluding VAT). The cost of MF was £2464, which was obtained from the NHS reference costs. 151 Cost of theatre, surgery for implantation of MACI/ACI was assumed to be the same as the cost of MF, though our clinical opinion is that MF usually requires an inpatient stay (because of pain), whereas ACI is usually a day case procedure. The Aastrom assumption may therefore slightly disadvantage MACI. The submission states that patients have one procedure. It is not clear whether this means that they would have only one MACI or it is an error by not accounting for both arthroscopy and harvesting, and later implantation. The manufacturers assumed that the cost of MACI/ACI reoperation would be £16,226. The cost of initial MF with MACI/ACI as second repair at an average cost of £17,623 also seems appropriate. This extra differential cost approximates to 3.5 extra rehabilitation visits. The cost of rehabilitation was £376, which was obtained from the Unit Costs of Health and Social Care (Curtis 2013154). This cost was based on the use of a community-based therapist, with 6–8 rehabilitation sessions, each lasting 30 minutes. Alternative rehabilitation costs could have been obtained from the NHS reference costs. 151 The budget impact model did not take into account any outpatient visits and any inpatient stays for MACI/ACI.

Three-year probabilities for MACI reoperation and MF reoperation were obtained from the SUMMIT trial data100 and these were converted to annual probabilities: 0.005 and 0.014, respectively. The annual probability for MACI was also assumed for the ACI reoperation, which seems a reasonable assumption. The Saris et al. 200970 data provided alternative 3-year probabilities for reoperation – these were converted to annual probabilities: 0.010 for ACI/MACI reoperation and 0.040 for MF reoperation. The manufacturers assumed that if MACI/ACI failed then a reoperation would be either MACI/ACI; however, if MF failed then a reoperation would be MACI/ACI.

The budget impact model explored two scenarios: one scenario with MACI/ACI as first-line treatment and the other scenario without MACI/ACI (with MF only). Using failure rates based on the SUMMIT data100 there were total cost savings from using MACI/ACI, ranging from approximately £5.9M in year 1 to £8.3M in year 5 – this was due to the lower reoperation rate and the expectation that 500 procedures (of the approximate 10,000 procedures) were either MACI/ACI. The Aastrom submission also included a budget impact model using the higher failure rates from Saris et al. 2009. 70 There were further total cost savings, although lower than the cost savings when the SUMMIT trial100 failure rates were used – using MACI/ACI the cost savings ranged from approximately £5.8M in year 1 to £7.8M in year 5 – these lower cost savings were due to the need for more reoperations after MF. In conclusion, the cost calculations provided by Aastrom seem reasonable and plausible.

The ACTIVE trial

The ACTIVE trial35 is a Medical Research Council-funded multicentre RCT of ACI against standard treatment, which could include debridement, abrasion, drilling, MF, mosaicplasty or bone graft (according to surgeon’s discretion) in 390 patients (195 in each group) with symptomatic chondral defect(s) on the medial or lateral femoral condyle, trochlea or patella, who had failed previous treatment and who were also considered suitable for ACI/MACI.

The protocol is available online from the ACTIVE trial35 website (www.active-trial.org.uk/).

Details of the results were made available to NICE for the ACI appraisal as academic in confidence. Only some of the results are reproduced here, with permission from the ACTIVE trial35 group.

Patients were recruited from 29 centres. The RJAH in Oswestry recruited 87 patients (22%). Six centres recruited between 20 and 29 patients, six centres recruited between 10 and 19 patients, and 16 centres recruited fewer than 10 patients.

Cost-effectiveness

The cost-effectiveness analysis by Oswestry was based on the ACTIVE trial35 data, but using cell costs only from Oswestry. The cells used in other ACTIVE35 sites came mainly from Genzyme. The submission provided the costs of ACI and the different comparators. The benefits were in terms of QALYs, which were estimated from the EQ-5D-3L.

The total costs and incremental costs with and without the market forces factor (MFF) provided in the submission have been summarised in Table 17 (MFF estimates the unavoidable cost differences of providing health care.) With Payment by Results (PbR), the MFF directly funds providers for the relative level of unavoidable costs they face. Each NHS Trust receives an individual MFF value used to establish the level of unavoidable costs they face relative to each other.

Accounting for unavoidable costs ensures a level basis across the country to provide equal amounts of health care per pound. So, in terms of PbR income, this would equal the activity multiplied by its tariff price, and this is then multiplied by the MFF value. All costs are in 2014–15 prices in UK pounds sterling (£). The second stage for ACI includes the cost of the cells. Production of cells in Oswestry costs £4125 per patient, but this does not include overheads and set-up costs. The submission stated that the incremental cost of ACI over TKR was £3746 and the incremental cost of ACI over MF, osteotomy or mosaicplasty was £7094.

Autologous chondrocyte implantation costs included operations, hospital stays, the cells and any further implants. The TKR cost is in line with NHS reference costs (2012–13), which is £5676 (NHS reference costs 2012–13151). The costs included only the direct costs of the procedures. No information in the submission was provided on any further outpatient or rehabilitation visits. The submission also stated that further data have been collected using patient diaries on patient and societal costs, such as any out-of-pocket expenses and time off work, but were not available in time for this submission.

In the ACTIVE trial,35 QALYs were estimated using the EQ-5D-3L version, weighted using the UK utility values. The control group consisted of patients who received MF, MF plus collagen membrane, and mosaicplasty. The ACI group had 192 observations at baseline, with a mean EQ-5D index score of 0.506, and by 8 years this score had increased to 0.676 (n = 29). The control group had 187 observations at baseline with a mean EQ-5D index score of 0.532 (slightly higher than the ACI group) and by 8 years this score was 0.647 (n = 29); this score was slightly lower than the corresponding score for the ACI group at 8 years. One year postoperatively, the EQ-5D index score had risen to 0.636 in the ACI group and 0.675 in the control group. The submission stated ‘these mean 1-year levels are largely maintained up to seven years post-surgery, after which the mean score for the control group seems to drop slightly’. Too few patients had reached the later years for a reliable analysis. Cumulative QALYs were also presented, which represented the total increase in health utility. These cumulative QALYs increased at a similar constant rate for both the ACI and control groups up to 6 years. The submission concluded that ‘at the 8-year time point, the ACI group has experienced – on average – an incremental gain of 1.19 QALYs more than the control group’ – this figure relates to cumulative QALYs and not mean QALYs.

Numbers of patients and EQ-5D for later years are reproduced in Table 18.

The submission concluded that the trial found that the ACI group gained more QALYs than the control group (incremental gain = 1.19 QALYs) at the 8-year time point, which resulted in a crude estimate of an incremental cost per QALY gain of £5961 for ACI over the control group.

In the base case, both groups were treated as homogeneous, but because of differences in the treatments for the control group and cell sources for the ACI group, further analyses tested for heterogeneity in each group. For the control groups (MF, MF plus collagen membrane and mosaicplasty), the data from the ACTIVE trial35 suggested little difference in the EQ-5D scores. For the ACI group, as the cells came from different sources, the submission included a regression analysis to see whether or not cell origin might affect their benefit. The regression analysis provided a negative value from which the authors concluded that ‘ACI patients treated outside Oswestry are unlikely to have more benefit from ACI’.

The absolute values for EQ-5D from the ACTIVE data35 often show little difference, as shown in Table 18, but changes from baseline EQ-5D show a more consistent advantage for ACI, as shown in Table 19.

Previous repairs

As reported earlier, in case series previous MF appears to reduce the success of ACI. The trials reviewed above do not contribute much evidence on this. Basad did not give details of previous surgery. In TIG/ACT43 only a few (8/57) of the ACI group had had previous MF. In the SUMMIT trial,100 32% of the MACI group had had previous repair attempts with MF, but this appeared to have little effect on response rates (no prior repairs, 90% response; more than one, 84%). In the ACTIVE trial,35 almost half had a previous repair procedure, but results are not given separately for them.

Several factors need to be considered in interpreting the evidence. First, we are somewhat reliant on subgroup analysis. Second, those who have had previous surgery may be older than those going straight to ACI, and chondrocyte viability declines with age. Third, some of the older trials had few patients who had not had prior surgery. Last, and most important, the evidence does not suggest that ACI is not worthwhile after prior MF, but only that it is not as successful. Hence there is no reason not to try ACI.

Duration of symptoms

In the SUMMIT trial,100 response rates were similar at 2 years: 82% for those with symptoms for less than 3 years, 92% in those with longer durations. Basad did not report results by prior duration but his MACI patients had a mean duration of symptoms of only 2.2 years. The ACTIVE trial35 did not report durations. The main evidence comes from the TIG/ACT43 5-year data for which only those with duration of symptoms under 3 years showed a significant difference between ACI and MF. Improvements in KOOS at 5 years were 26 for the ACI group versus 15 for the MF group (p = 0.026). For the subgroup with over 3 years’ duration, KOOS improvements were 13 for ACI and 17 for MF (not significant). This might suggest that ACI is of less value, relative to MF, in patients with longer duration.

Previous studies have shown improvements with ACI after long duration of symptoms. In the trial by Bentley et al. ,78 most patients receiving ACI had excellent Cincinnati scores results despite a mean duration of symptoms of 7.2 years. In the trial of ACI-C versus MACI by Bartlett et al. ,57 59% of the ACI-C group and 72% of the MACI group had good or excellent Cincinnati scores despite duration of symptoms of approximately 10 and 7 years, respectively. In another study from Stanmore by Biant et al. ,79 of a cohort of 104 ACI patients followed for at least 10 years, 66% had excellent or good Cincinnati scores despite an average duration of symptoms before ACI of 7.8 years.

ACI-C or MACI?

In a very large cohort of 827 patients with mean duration of follow-up 6.2 years, Nawaz et al. 80 reported better results with MACI than ACI-P or ACI-C, though this was probably due to different durations, as the ACI groups came from an earlier period and so had more time for knee status to decline. The RCT of ACI-C versus MACI by Bartlett et al. 57 found no difference at 1 year.

In practice, ACI has evolved and most use is now expected to be MACI, with characterised chondrocytes.

Predicting success

Nawaz et al. 80 noted that the chance of success was reduced by previous attempts at repair and the presence of any osteoarthritic change in the knee. They summarised the results from their very large cohort study thus:

Our study suggests that the ‘ideal’ candidate for autologous chondrocyte implantation is a younger individual with a single lesion on the trochlea or the lateral femoral condyle, with no previous procedures or evidence of degenerative changes.

At 12 years after ACI, their ideal patients had a repair survival of almost 80% compared with 50% in the whole group. Graft survival was < 25% in those who had had a previous repair (including MF and mosaicplasty) and > 75% in those who had not.

Survival of repairs

The 2-year differences between MF and ACI or MACI arise mainly because symptom scores reach a plateau sooner after MF than after ACI. Saris et al. 70 reported (from graph) that a KOOS plateau was reached with MF by 12–18 months, whereas improvements continued after ChondroCelect ACI-P. The SUMMIT100 investigators showed a plateau before 12 months with MF but not until 18 months with MACI.

In the TIG/ACT trial,43 Saris et al. 70 reported (from graph, so approximate) that about 7% of MF repairs had failed by 20 months and 11.5% by 36 months (but based on only seven failures in the MF group). The longer-term results reported by Vanlauwe et al. 43 showed the plateau in the KOOS in the MF group from 12 to 60 months, whereas the ChondroCelect group with duration of symptoms < 3 years at surgery, reached a plateau at 36 months. The CC group with duration of symptoms of > 3 years showed no difference in KOOS from MF with an early plateau and lines almost overlapping.

Basad et al. 64 reported that the Lysholm score in the MF group improved from 55 at baseline to 81 at 12 months, but then declined to 69 at 24 months. The MACI group had a baseline score of 55, improving to 95 at 12 months, maintained at 92 at 24 months.

Bhosale et al. ,108 from Oswestry, report results at an average of 5 years (range about 3–9 years) among 80 patients, all but four having had ACI-P. The median baseline Lysholm score was 54, which improved to a median of 78 at 12 months post operation. Of the 80 patients, 65 improved and scores at 15 months were maintained for up to 9 years. They also reported that higher age, female gender and larger defect size were associated with greater benefit, but none of these associations was statistically significant. They concluded that a good result at 15 months is durable.

Clinical pathways

Figure 21 shows the detailed clinical pathway for people receiving treatment for symptomatic articular cartilage defects of the knee.

FIGURE 21.

Clinical pathways for patients with articular cartilage defects of the knee joint.

Knee repairs

The starting point for the model is the primary repair, which could be either ACI or MF. After the primary repairs, patients can then either move to the ‘successful primary repair’ health state or to the ‘failure of primary repair’ health state. Success can be permanent: the first repair works and they do not require a second repair. So they can stay in the ‘successful primary repair’ health state until they die. Or success can be temporary (the patient has no symptoms for years but after a while the repair fails), so the patient then moves to the ‘failure of primary repair’ health state. They can then have a second repair or they can choose not to have another repair (no further repair health state) and rely on analgesics to relieve symptoms, that is, the patient chooses to accept the pain and treat it rather than have another attempt at repair, though later he/she may have a knee replacement.

The second repair could be either ACI or MF. Based on clinical opinion, we have assumed that patients will have a maximum of two repairs. Once the patient has had a second repair he/she can then either move to the ‘successful second repair’ health state or to the ‘failure of second repair’ health state. The successful second repair can be permanent (similar to the successful primary repair) and patients stay in this health state until they die. Or it could be a temporary success, so then the patient later moves to the ‘failure of second repair’ health state. We are assuming that patients whose second repair fails do not have another repair and they move to the ‘no further repair’ health state.

Patients who move to the ‘no further repair’ health state after failure of repair can choose not to have another repair procedure and accept the pain, taking analgesics as required (that is, they can stay in this health state), until they reach the knee replacement age range, when their options are knee replacement or continued symptomatic treatment. Those who choose not to have a further repair may have had partial relief from symptoms, so we rate their utility as better than the baseline one.

Knee replacements

We assume for simplicity that patients over the age of 55 years cannot have an ACI or MF, but have only a knee replacement or symptomatic care. This is line with the MSAC report, which indicated that MACI/ACI was not indicated for patients older than 55 years. 32 The first knee replacement can be either a partial (unicompartmental) knee replacement (PKR) or TKR. According to statistics from the National Joint Registry, the average ages of patients having a PKR and TKR are 64 and 70 years of age, respectively. 157 However, we know that patients being considered for ACI are a lot younger than the general population (average age early thirties), so that if the repair fails, they are more likely to have a knee replacement at an earlier age. In line with expert clinical advice, we are assuming that patients can have one or more knee replacements. The first may succeed for life; if not, they can have another replacement or choose not to have another replacement. The first knee replacement could be either a PKR or a TKR, but we have assumed that all subsequent replacements will be TKRs.

A patient can move to ‘first knee replacement’ from either a temporarily ‘successful primary repair’ health state, a temporarily ‘successful second repair’ health state or from the ‘no further repair’ health state when they reach the knee replacement age range. The first knee replacement can be a success, so the patient moves to the ‘successful first knee replacement’ health state or the replacement can fail over time, so they move to the ‘failure of first knee replacement’ health state. The first knee replacement can be a permanent success until the patient dies or a temporary success because the knee replacement fails over time, so he/she moves to the ‘failure of first knee replacement’ health state, from which patients can choose to have another knee replacement or to have no further knee replacement (so move to the ‘no further knee replacement’ health state).

The second knee replacement can be a permanent success (until death) or it could be a temporary success, and patients move to the ‘failure of further knee replacement’ health state, from which they can choose to have no further knee replacement (but use symptomatic treatment) and or to have another (third) knee replacement. Based on clinical opinion, we have assumed that patients can have more than two knee replacements. Patients who move to the ‘no further knee replacement’ health state choose not to have another knee replacement and stay in this health state until they die.

Deaths

Patients can move to death from any of the repair and replacement health states due to all-cause mortality, or because of the rare mortality associated with PKR or TKR (such as DVT and pulmonary embolism). The latter becomes more relevant in later stages because replacing previous knee replacements requires more extensive procedures.

Markov model structure

The Markov model structure is shown in Figure 22. In line with the clinical pathway shown in Figure 21, the model shows the different health states and events that can take place. The different events health states for the model are shown by the ovals. The model shows all of the transitions that can happen between the different health states by the direction of the arrows. The little loop arrows in the left-hand corner of the ovals (recurring arrow) means that a patient can stay in that health state for more than one cycle, and perhaps indefinitely (until they die). The dashed line indicates that at 55 years of age, the patient can choose to have a knee replacement (total or partial). Transition probabilities, that is, the rate of progression from one health state to another (or for staying in the same health state), were identified from the literature.

FIGURE 22.

Markov model structure for patients with articular cartilage defects of the knee joint.

Base-case analysis

Many people with cartilage injury are young and involved in sports, and this is where most of the injuries occur. We have not differentiated by gender, as evidence shows that there is no difference in the success or failure of the two different procedures (ACI and MF) if lesions are comparable. 146 For the base-case analysis, we have adopted a lifetime horizon (i.e. patients can live to 100 years), with a cycle length for the model set at 1 year, and transitions between each health state occurring at the end of each cycle. A cycle length of 1 year is reasonable, given the time that it takes patients to recover from surgery. A hypothetical cohort of a 1000 patients with symptomatic articular cartilage defects of the knee with a starting age of 33 years is followed from their first repair. The analysis is conducted from the perspective of the NHS and PSS. All costs are in UK pounds sterling in 2012–13 prices. Health outcomes were measured in QALYs. Results are expressed as incremental cost per QALY gained. An annual discount rate of 3.5% is applied to both costs and outcomes.

Transition probabilities

For the base-case analysis, annual transition probabilities were based on data derived from the literature and in consultation with clinical experts. For the primary and second repairs for both ACI and MF, these transition probabilities were based on success rates for ACI compared with MF, and these probabilities came from two main studies: Saris et al. 200970 and 2014. 100

Figure 23 shows a flow chart with the proportion of people achieving success or failure with each repair. Saris et al. 70 reported a success rate of 83.0% over a 3-year period for ACI, and, of the remaining 17% of people, 3.9% would require reoperation of the same lesion (either ACI or MF as a second repair). We therefore assumed that 13.1% of people would have no further repair following the first repair. After ACI as a first repair, for those patients who required a second repair, and if this repair was another ACI, we assumed the same success rate for this second repair as the first repair. If this second repair was MF, we used information from Vanlauwe et al. ,43 who reported that for 16.4% of people MF after a first ACI failed at 5 years. We assumed that the rest of the people had successful MF procedures following an ACI.

FIGURE 23.

Proportion of patients achieving success/failure with ACI or MF at 36 months.

Saris et al. 70 also reported a success rate of 62.0% over a 3-year period for MF and, of the remaining 38% of people, 11.5% would require reoperation of the same lesion (either ACI or MF as a second repair). We therefore assumed that 26.5% of people would have no further repair following the first repair. After MF as a first repair, for those patients who required a second repair and if this repair was MF, we assumed the same success rate for this second repair as the first repair. If this second repair was an ACI, this information was obtained from Biant et al. ,79 who reported that for 30.9% of people an ACI after a first MF failed at 10 years. We assumed that the rest of the people had a successful ACI procedure following MF.

The timing of knee replacement after one of the repair health states was based on data from the RCT of ACI and MF by Knutsen et al. 67 Transition probabilities for success and failure for patients who needed knee replacements or knee replacement revisions were derived from two studies: Gerlier et al. 141 and Dong and Buxton. 158 Appendix 11 details the literature used and the assumptions made for deriving these probabilities, and Tables 57–59 show the transition probabilities that have been used in the base-case analysis.

Utilities

There are very few studies reporting health-state utility values for patients with symptomatic articular cartilage defects of the knee. The main studies reporting utility values have been summarised in Chapter 4 (Clar 2005,3 Derrett 2005,140 Gerlier 2010141). In the previous HTA report, the preoperative QoL value was taken to be 0.80 and for those who had successful knee repair there was a utility gain of 0.10 (utility score for successful knee repair was 0.90); for those where the knee repair failed the utility value remained at the preoperative value (utility score for knee repair failure was 0.80) (Clar 20053). Derrett et al. 140 used the EQ-5D-3L to elicit utility scores, and the ACI waiting list group had a pre-surgery utility score of 0.41. After surgery, the EQ-5D-3L mean score for the ACI group was 0.64 and for mosaicplasty was 0.47, a utility gain of 0.23 and 0.06, respectively.

For our model we have used utility values for knee repairs from the Gerlier et al. study,141 comparing ACI with MF using data from the TIG/ACT43 ChondroCelect trial. They used a short-term model with a time horizon of 5 years to take into account knee pain and mobility after the initial intervention (QoL information was obtained from a 5-year RCT using the SF-36) and also a long-term model with a time horizon of 40 years to take into account the development of OA after 15 years and the need for a TKR after 20 years. We used two other studies to supplement utility values for knee replacement. The first study is by Dong and Buxton,158 who developed a Markov model to compare the cost-effectiveness of TKR using computer-assisted surgery with that of TKR using a conventional manual method in the absence of formal clinical trial evidence. The second study is by Jansson and Granath,159 who analysed EQ-5D data before and after knee arthroplasty.

Table 21 shows the base-case mean utility values used in the model. For the repairs these values were all obtained from the paper by Gerlier et al. ,141 who used the SF-36 and the KOOS measures to estimate utility scores. The mean utility value for patients before they have a primary repair (before ACI or MF) was 0.654: this utility value was based on the initial value before the intervention. For those patients who had an ACI as a first repair and moved to the ‘successful primary repair’ health state, we assumed that the patients’ mean utility value after surgery for the first year would be 0.760 (this value was based on year 1 post intervention, regardless of the outcome, and takes into account the long rehabilitation period and abstinence from active pursuits), and, if they remain in this health state in subsequent years, their utility value would remain constant at 0.817. This latter value was based on patients who had clinical success for 5 years after the intervention. For those patients who had MF as a first repair and moved to the ‘successful primary repair’ health state, we assumed that the patients’ mean utility value after surgery for the first year would be 0.760 (this value was based on year 1 post intervention, regardless of the outcome). For years 2–4 after MF this mean utility value would increase to 0.817. This reflects the quite long rehabilitation required in the first year after the procedures, and the time taken for the cartilage to be replaced. For years 5 and onwards for patients who stay in this same health state, we have assumed that utility would fall to the same as that pre-surgery (mean utility value is 0.654) because the benefit of MF declines after 5 years and patients may choose to have another procedure.

For those patients who require a second repair, the mean utility value was 0.654: this value was based on the utility value before the intervention. 141 For those requiring a second repair there are four possible sequences: ACI(ACI), ACI(MF), MF(ACI) and MF(MF).

Utilities for patients having a second successful ACI after the first ACI were assumed to be the same as for those who had a successful ACI as a first repair. Utilities for patients having a successful MF after an initial ACI that failed were assumed to be the same as for those who had a successful MF as a first repair and moved to the ‘successful primary repair’ health state.

Patients who have a successful ACI after an initial MF move to the ‘successful second repair’ health state. However, as noted previously, ACI is less effective in patients who have had prior MF, so for years 4 and 5 we have used the average of two utility values from Gerlier et al. ,141 based on year 1 post intervention (utility value = 0.760) and clinical success after 5 years following the intervention (utility value = 0.817), so the mean utility value for ACI after MF was 0.789. Utilities for patients having a successful second MF after a failed initial MF and who moved to the ‘successful second repair’ health state were assumed to be the same as those who had a successful MF as a first repair.

For patients who moved to the ‘no further repair’ health state, the mean utility value was 0.691: this value was based on patients who had not had a successful result 5 years after surgery (Gerlier 2010141), but we have assumed that those who choose to have no further repair may have had some benefit from the first repair, and so do not go back as far as the original baseline utility.

Mean utility values are the same for knee replacements after ACI or MF. Before the first knee replacement procedure, patients who received a TKR and PKR are assumed to have the same utility value, 0.615. This value was based on an average of two utility values: (1) the EQ-5D index score at baseline preoperatively for knee arthroplasty (value = 0.51)159 and (2) an estimated value for TKR operation for knee problem (value = 0.72). 158 For patients who move to the ‘successful first knee replacement’ health state (TKR or PKR), a utility value of 0.780 was also obtained from Dong and Buxton. 158 This utility value was estimated from the generic Knee Society Score scale and was applied to the Markov health state for normal health after primary TKR. We have also assumed that if patients move to the successful ‘further total knee replacement’ health state then they will have the same utility value as if it was a first TKR. For those patients for whom TKR has failed and need a further TKR, the utility value was 0.557 based on the failed TKR/revision health state from Gerlier et al. 141 Finally, for those patients who move to the ‘no further replacement health state’ this value (mean = 0.691) was also from Gerlier et al. 141 and was based on patients who had no clinical success 5 years after surgery (in line with patients who move to the ‘no further repair’ health state).

Resource use and costs

Costs for the different procedures (ACI, MF, PKR/TKR, TKR revisions) and for outpatient visits and rehabilitation are shown in Table 22. We have used national reference costs where possible (NHS reference costs 2013151) supplemented by the previous HTA report on ACI 3. All unit costs are presented in pounds sterling (£) in 2012–13 prices.

The cost of the ACI (ChondroCelect and MACI) includes the costs associated with cell development, including the ACI kit, staff time and transporting the cells to and from the laboratory. ACI involves two procedures: the arthroscopic cell harvest and the re-implantation during arthrotomy. We assumed both would be done as day cases. Based on clinical experience, we have also included the costs of six outpatient visits and three rehabilitation visits in the first year (Table 23).

The cost of MF procedure (including an inpatient stay) was obtained from Clar et al. 3 and the cost has been updated to 2012–13 prices using the Hospital and Community Health Services (HCHS) index. 154 The inpatient stay is required because, unlike after ACI, the patients can have considerable pain after MF because of the drilling into bone. Over the course of the year, the patient would also have three outpatient visits and three rehabilitation visits, and these costs have been added for this health state (based on clinical experience).

The cost for a first knee replacement was obtained from the NHS reference costs151 and we have assumed that it could be either a TKR or a PKR. After a TKR, a subsequent TKR is almost double the cost, because it is technically more difficult. After a PKR, a second knee replacement would be a TKR, and we have assumed that this would cost £5676. If the patient required any more subsequent knee replacements (all of which would be TKRs) then these would cost £12,959. Based on consultation with clinical experts, in the first year after knee replacement, we have included the cost of two outpatient visits (see Table 23).

Unit costs were obtained from the NHS reference costs. 151

We have assumed that there will be no further costs after year 1 once patients enter the successful health states (‘successful primary repair’, ‘successful second repair’, ‘successful first knee replacement’ and ‘successful further knee replacement’), as patients incur costs such as outpatient or rehabilitation visits during the first year of either a knee repair or a knee replacement. In addition, for the ‘no further repair’ health state or the ‘no further knee replacement’ health state, we have not added any costs for the analgesics based on advice from our clinical experts, as these costs are negligible and these patients are not followed up routinely, and it is up to the GP to refer the patient back to the hospital for a knee repair or a knee replacement.

Complications

Adverse events have not been included, as there were no important differences between the two treatment arms.

Mortality

Age-specific mortality rates used in the economic model were based on the UK general population lifetime tables from the ONS. 160 Using the ONS data, the average probabilities of death for men and women were combined. As the cohort ages, mortality rates generally increase throughout the time horizon in the model. In the model, patients from any health state can move to the ‘dead’ health state. Patients undergoing a knee replacement are subject to a mortality risk during surgery. To reflect this higher mortality, rates were obtained from a study by Mahomed et al. 150 For those patients undergoing a TKR and a TKR revision, the mortality rates were reported as 0.7% and 1.1%, respectively.

Measuring cost-effectiveness

The base-case analysis assessed the cost-effectiveness of ACI compared with MF. We calculated for a cohort of patients the expected quality-adjusted survival based on their likelihood of surviving each cycle, their expected health-state utility value, and their expected costs. We have adopted a lifetime horizon from a starting age of 33 years. The analysis is conducted from the perspective of the UK NHS and PSS. Costs are expressed in 2012–13 UK pounds sterling. The main outcome of interest was the QALY. The different sequences of procedures were ranked in order of increasing cost. We eliminated any categories for which another category was cheaper and more effective (simple dominance). If the ICER for a given category was higher than that of the next more effective alternative then this category was eliminated (extended dominance). For the remaining options, we reported the ICERs, measured as cost per QALY gained. Discount rates of 3.5% were applied to both future costs and benefits, as costs and benefits accrued in the future are valued at less than those accrued today.

Sensitivity analysis assesses the uncertainty in parameter inputs used in the Markov model and to check whether or not the results obtained are robust. We present both deterministic and probabilistic results. For the deterministic analysis, we identified the key factors driving the cost-effectiveness. For the PSA, to reflect the amount and pattern of the variation, the analysis attributes probability distributions randomly around specified parameters with simulations, which are repeated to generate ICERs. The PSA was undertaken using 1000 simulations. We used the gamma distribution for costs and the beta distribution for utility values and transition probabilities. 161 As the values for costs, utilities and transition probabilities used in the model were means or weighted averages, an assumption was made for the SE in order to calculate the alpha and beta values, which are required for the PSA. For example, we have assumed the SE to be 0.1 of the mean value. 162,163 These bootstrapped simulations obtained from the PSA were used to construct cost-effectiveness acceptability curves (CEACs) to illustrate the effect of sampling uncertainty in which individual model parameters were sampled from the appropriate probability distribution. CEACs were presented to indicate the probability of a procedure being cost-effective using a willingness-to-pay threshold from £0 to £60,000.

Base-case cost-effectiveness results

One thousand patients entered the model, with a starting age of 33 years. For the primary repair these patients can receive either an ACI or MF, and if these patients require a second repair it could either be an ACI or MF. Many will not require a second repair, but the cost-effectiveness of the primary repair depends partly on the costs of subsequent interventions required or avoided, so we need to consider the sequence options.

Table 24 shows the base-case deterministic and probabilistic cost-effectiveness results for the lifetime horizon for the two different scenarios. For scenario 1, if patients required a second repair this would be an ACI and for scenario 2 if patients required a second repair this would be MF. After MF, 11.9% of patients required a second procedure, and after ACI 3.9% of patients required a second repair. Looking at the discounted deterministic results, for scenario 1, ACI costs £14,314 more than MF, but generated 0.9944 more QALYs than MF. The cost per QALY gained for ACI compared with MF was £14,395. For scenario 2, ACI, again, was more costly (incremental cost = £14,877), but generated more QALYs (0.9537) and the resulting cost per QALY gained was £15,598. For both scenarios, ACI as a first repair was more cost-effective than MF as a first repair. These results were of similar magnitudes and directions for both the undiscounted deterministic results and the probabilistic results.

One of the key cost drivers was the cost of the cells for the ACI procedure, but, over the lifetime horizon, there are QALYs gained from using ACI, and there are cost savings to the NHS later due to fewer people needing a second repair, fewer people in need of a TKR, and fewer people moving to the no further repair/replacement health states (in which the utility is lower).

Figure 24 presents the CEACs for the base-case results for scenarios 1 and 2, respectively. For scenario 1, if the decision-maker was willing to pay £14,000, the probability that both ACI and MF were cost-effective was approximately 50%; however, if the decision-maker was willing to pay £20,000, ACI was probably 59% more likely to be cost-effective than MF (see Figure 24a). These results were similar for scenario 2: if the decision-maker was willing to pay £20,000, the probability that ACI was more cost-effective than MF was 56% (see Figure 24b).

FIGURE 24.

Cost-effectiveness acceptability curve: base case results: (a) scenario 1; and (b) scenario 2.

Table 25 shows the base-case deterministic and probabilistic cost-effectiveness results for the lifetime horizon, ranked by the least costly sequence (option). For the discounted deterministic results MF(MF) was the least costly option and had the fewest QALYs, whereas ACI(ACI) was the most expensive option but generated more QALYs. The option MF(ACI) was extendedly dominated by a linear combination of MF(MF) and ACI(MF) and therefore this option was eliminated from the comparison. The ICER ACI(MF) and MF(MF) was just under £15,000; doing ACI first is more cost-effective. The ICER between the two initial ACI options: ACI(ACI) vs. ACI(MF) was just under £16,000; even if the first ACI fails, there is good enough chance of a second ACI succeeding to make the ICER for a repeat ACI quite reasonable. So initial ACI appears more cost-effective than initial MF and for those who need a second repair after the first ACI, this should be another ACI. These results were of similar magnitudes and directions for the probabilistic results and also for the undiscounted deterministic results.

Figure 25 presents the CEAC for the base-case results for all sequences. The graph shows that for amounts below £14,000 then MF(MF) appears cost-effective compared with the other three options. At a willingness to pay of £16,000, there is not much difference between the four options. However, if the decision-maker is willing to pay £18,000 or more for a QALY then ACI as a first procedure [either ACI(ACI) or ACI(MF)] is probably more cost-effective than MF [either MF(ACI) or MF(MF)] as a first procedure.

FIGURE 25.

Cost-effectiveness acceptability curve: base-case results – all sequences.

Scenario and sensitivity analysis cost-effectiveness results

This section highlights the results from the different SAs that were undertaken.

Sensitivity analysis 1: cell cost reduction

In the base-case analysis, the cost of cells for ChondroCelect and MACI procedures was taken to be £16,000. We know that there are confidential discounts provided to the NHS by manufacturers. In this SA we have varied this figure by reducing the cell costs by 25% (£12,000), 50% (£8000) and 75% (£4000). The last figure may seem very low, but it is similar to the cost provided in the Oswestry submission for cells produced in a NHS facility.

Table 26 shows the results when the cost of cells is reduced. When there was a reduction in cell costs, even though ACI was more costly than MF, there were more QALYs gained with ACI than MF. For a 25% cell cost reduction, the deterministic cost per QALY gain ratio for ACI compared with MF was £10,523 for scenario 1 and £11,404 for scenario 2. The cost per QALY gain ratio for a 50% cell cost reduction for ACI compared with MF was £6651 (scenario 1) and £7210 (scenario 2), and the resulting figures for a 75% reduction were £2779 (scenario 1) and £3016 (scenario 2). With the reduction in cell costs, the cost-effectiveness of ACI improved relative to MF. Hence, the cost of cells was a key driver for the cost-effectiveness. These results were of similar magnitudes and directions for the probabilistic results.

TABLE 26 - Sensitivity cost-effectiveness results (by scenario): cell cost reduction

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic: 25% reduction
Scenario 1
MF(ACI)	6183	17.0284	–	–	–
ACI(ACI)	16,647	18.0228	10,464	0.9944	10,523
Scenario 2
MF(MF)	5015	17.0033	–	–	–
ACI(MF)	15,892	17.9570	10,877	0.9537	11,404
Probabilistic: 25% reduction
Scenario 1
MF(ACI)	6183	17.0305	–	–	–
ACI(ACI)	16,637	18.0497	10,454	1.0192	10,258
Scenario 2
MF(MF)	5009	17.0086	–	–	–
ACI(MF)	15,880	17.9502	10,871	0.9416	11,545
Deterministic: 50% reduction
Scenario 1
MF(ACI)	5760	17.0284	–	–	–
ACI(ACI)	12,373	18.0228	6614	0.9944	6651
Scenario 2
MF(MF)	5015	17.0033	–	–	–
ACI(MF)	11,892	17.9570	6877	0.9537	7210
Probabilistic: 50% reduction
Scenario 1
MF(ACI)	5770	17.0250	–	–	–
ACI(ACI)	12,362	18.0100	6592	0.9850	6693
Scenario 2
MF(MF)	5020	16.9907	–	–	–
ACI(MF)	11,876	17.9123	6856	0.9216	7439
Deterministic: 75% reduction
Scenario 1
MF(ACI)	5336	17.0284	–	–	–
ACI(ACI)	8100	18.0228	2763	0.9944	2779
Scenario 2
MF(MF)	5015	17.0033	–	–	–
ACI(MF)	7892	17.9570	2877	0.9537	3016
Probabilistic: 75% reduction
Scenario 1
MF(ACI)	5346	16.9755	–	–	–
ACI(ACI)	8083	18.0442	2737	1.0687	2561
Scenario 2
MF(MF)	5023	16.9546	–	–	–
ACI(MF)	7878	17.9253	2854	0.9707	2940

Figure 26a–f presents the CEACs for the SA for cell cost reductions for scenarios 1 and 2. For a 25% cell cost reduction: scenario 1, if the decision-maker was willing to pay £20,000, the probability that ACI was more cost-effective than MF was 65% (see Figure 26a); scenario 2, the probability that ACI was more cost-effective than MF was 58% (see Figure 26b). For a 50% cell cost reduction: scenario 1, if the decision-maker was willing to pay £20,000, the probability that ACI was more cost-effective than MF was 67% (see Figure 26c); scenario 2, the probability that ACI was more cost-effective than MF was 64% (see Figure 26d). For a 75% cell cost reduction: for scenario 1, if the decision-maker was willing to pay £20,000, there was a 71% probability that ACI was more cost-effective than MF (see Figure 26e); scenario 2, there was a 70% probability that ACI was more cost-effective than MF (see Figure 26f). The graphs indicate that reductions in cell costs improve the cost-effectiveness of ACI compared with MF.

FIGURE 26.

Cost-effectiveness acceptability curves: cost reductions: (a) 25% cell cost reduction – scenario 1; (b) 25% cell cost reduction – scenario 2; (c) 50% cell cost reduction – scenario 1; (d) 50% cell cost reduction – scenario 2; (e) 75% cell cost reduction – scenario 1; and (f) 75% cell cost reduction – scenario 2.

Table 27 shows the deterministic and probabilistic cost-effectiveness results for the lifetime horizon for cell cost reduction, and results were ranked by least costly option. When the cost of cells was reduced by 25%, 50% and 75%, these results were in line with the base-case cost-effectiveness results. That is, for the discounted deterministic results MF(MF) was the least costly option and had the fewest QALYs, whereas ACI(ACI) was the most expensive option, but generated more QALYs. Again, the option MF(ACI) was extendedly dominated by a linear combination of MF(MF) and ACI(MF) and therefore this option was eliminated from the comparison. The deterministic ICER between ACI(MF) and MF(MF) was just over £11,000; and the ICER between the two ACI options: ACI(ACI) vs. ACI(MF) was just under £11,500 when there was a 25% reduction in costs. These ICER figures are £7200 and £7300, respectively, when there was a 50% reduction in costs, and the corresponding figures are £3000 and £3200, respectively, when there was a 75% reduction in costs. For all cell cost reduction scenarios, both the deterministic and probabilistic results indicate that ACI as a first procedure was more cost-effective than MF as a first procedure, and from these results, again, we see that the cost of cells is a key driver of the cost-effectiveness estimates.

TABLE 27 - Sensitivity cost-effectiveness results: cell cost reduction

As the incremental QALYs are near zero, the ICER can fluctuate widely.

Procedure	Total mean costs (£)	Total mean QALYs	Comparison	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic: 25% reduction
MF(MF)	5015	17.0033	–	–	–	–
MF(ACI)	6183	17.0284	MF(ACI) vs. MF(MF)	1168	0.0251a	Extendedly dominated
ACI(MF)	15,892	17.9570	ACI(MF) vs. MF(MF)	10,877	0.9537	11,404
ACI(ACI)	16,647	18.0228	ACI(ACI) vs. ACI(MF)	755	0.0658a	11,483
Probabilistic: 25% reduction
MF(MF)	5009	17.0086	–	–	–	–
MF(ACI)	6183	17.0305	MF(ACI) vs. MF(MF)	1174	0.0219a	Extendedly dominated
ACI(MF)	15,880	17.9502	ACI(MF) vs. MF(MF)	10,871	0.9416	11,545
ACI(ACI)	16,637	18.0497	ACI(ACI) vs. ACI(MF)	758	0.0994a	7618
Deterministic: 50% reduction
MF(MF)	5015	17.0033	–	–	–	–
MF(ACI)	5760	17.0284	MF(ACI) vs. MF(MF)	744	0.0251a	Extendedly dominated
ACI(MF)	11,892	17.9570	ACI(MF) vs. MF(MF)	6877	0.9537	7210
ACI(ACI)	12,373	18.0228	ACI(ACI) vs. ACI(MF)	481	0.0658a	7319
Probabilistic: 50% reduction
MF(MF)	5020	16.9907	–	–	–	–
MF(ACI)	5770	17.0250	MF(ACI) vs. MF(MF)	750	0.0343a	Extendedly dominated
ACI(MF)	11,876	17.9123	ACI(MF) vs. MF(MF)	6856	0.9216	7439
ACI(ACI)	12,362	18.0100	ACI(ACI) vs. ACI(MF)	486	0.0977a	4979
Deterministic: 75% reduction
MF(MF)	5015	17.0033	–	–	–	–
MF(ACI)	5336	17.0284	MF(ACI) vs. MF(MF)	321	0.0251a	Extendedly dominated
ACI(MF)	7892	17.9570	ACI(MF) vs. MF(MF)	2877	0.9537	3016
ACI(ACI)	8100	18.0228	ACI(ACI) vs. ACI(MF)	207	0.0658a	3155
Probabilistic: 75% reduction
MF(MF)	5023	16.9546	–	–	–	–
MF(ACI)	5346	16.9755	MF(ACI) vs. MF(MF)	322	0.0209a	Extendedly dominated
ACI(MF)	7878	17.9253	ACI(MF) vs. MF(MF)	2854	0.9707	2940
ACI(ACI)	8083	18.0442	ACI(ACI) vs. ACI(MF)	205	0.1189a	1725

Figure 27a–c presents the CEACs for the SA results for the cell cost reduction. The graphs clearly show that if the decision-maker is willing to pay £20,000 then the probability that ACI(ACI) is more cost-effective than the other three comparisons is 32% for a 50% reduction in the costs of cells (although there is not much difference if MF was the second repair after the ACI) and 38% for a 75% reduction in the costs of cells. Whereas, if the decision-maker pays £30,000 for a 25% reduction in the cost of cells then the probability that ACI(ACI) is more cost-effective than the other three comparisons is 34%. This suggests that ACI as first procedure is more cost-effective than MF as first procedure.

FIGURE 27.

Cost-effectiveness acceptability curves: cell cost reduction: (a) 25% cell cost reduction; (b) 50% cell cost reduction; and (c) 75% cell cost reduction.

Sensitivity analysis 2: changing the time horizon

In the base-case analysis, a lifetime horizon was chosen with the starting age 33 years for the cohort. In this SA we have varied the time horizon (10, 20, 30, 40 and 50 years) to see how this affects the ICER. Table 28 shows the sensitivity cost-effectiveness results for the different time horizons. For all of the time horizons, even though ACI was more costly than MF, there were more QALYs gained with ACI than MF. For the 10-year time horizon, the deterministic cost per QALY gain for ACI compared with MF was £25,992 for scenario 1 and £27,388 for scenario 2. The cost per QALY gained for the two scenarios ranged from £17,000 to £18,000 for the 20-year time horizon; £15,000 to £16,000 for the 30- and 40-year time horizons; and £14–16,000 for the 50-year time horizon. For both scenarios, ACI as a first repair was more-cost-effective than MF as a first repair, the longer the time horizon. These results were of similar magnitudes and directions for the probabilistic results.

TABLE 28 - Sensitivity cost-effectiveness results (by scenario): changing the time horizon

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic: 10-year time horizon
Scenario 1
MF(ACI)	5983	7.3030	–	–	–
ACI(ACI)	20,082	7.8454	14,098	0.5424	25,992
Scenario 2
MF(MF)	4498	7.2906	–	–	–
ACI(MF)	19,329	7.8321	14,831	0.5415	27,388
Probabilistic: 10-year time horizon
Scenario 1
MF(ACI)	5989	7.2950	–	–	–
ACI(ACI)	20,075	7.8501	14,086	0.5550	25,379
Scenario 2
MF(MF)	4505	7.2845	–	–	–
ACI(MF)	19,326	7.8270	14,821	0.5425	27,320
Deterministic: 20-year time horizon
Scenario 1
MF(ACI)	6104	11.2812	–	–	–
ACI(ACI)	20,340	12.1040	14,286	0.8228	17,301
Scenario 2
MF(MF)	4524	11.2587	–	–	–
ACI(MF)	19,384	12.0654	14,860	0.8067	18,421
Probabilistic: 20-year time horizon
Scenario 1
MF(ACI)	6098	11.2630	–	–	–
ACI(ACI)	20,359	12.1136	14,261	0.8506	16,766
Scenario 2
MF(MF)	4526	11.2362	–	–	–
ACI(MF)	19,414	12.0625	14,888	0.8263	18,019
Deterministic: 30-year time horizon
Scenario 1
MF(ACI)	6329	13.9997	–	–	–
ACI(ACI)	20,614	14.9318	14,285	0.9321	15,326
Scenario 2
MF(MF)	4739	13.9750	–	–	–
ACI(MF)	19,609	14.8774	14,871	0.9024	16,480
Probabilistic: 30-year time horizon
Scenario 1
MF(ACI)	6326	14.0117	–	–	–
ACI(ACI)	20,642	14.9494	14,316	0.9377	15,267
Scenario 2
MF(MF)	4728	13.9748	–	–	–
ACI(MF)	19,628	14.8754	14,900	0.9006	16,545
Deterministic: 40-year time horizon
Scenario 1
MF(ACI)	6492	15.7604	–	–	–
ACI(ACI)	20,798	16.7368	14,306	0.9764	14,652
Scenario 2
MF(MF)	4901	15.7354	–	–	–
ACI(MF)	19,775	16.6747	14,875	0.9393	15,836
Probabilistic: 40-year time horizon
Scenario 1
MF(ACI)	6494	15.7558	–	–	–
ACI(ACI)	20,763	16.7219	14,269	0.9662	14,768
Scenario 2
MF(MF)	4900	15.7279	–	–	–
ACI(MF)	19,376	16.6504	14,837	0.9225	16,083
Deterministic: 50-year time horizon
Scenario 1
MF(ACI)	6579	16.7164	–	–	–
ACI(ACI)	20,891	17.7078	14,313	0.9914	14,437
Scenario 2
MF(MF)	4987	16.6913	–	–	–
ACI(MF)	19,864	17.6427	14,876	0.9514	15,636
Probabilistic: 50-year time horizon
Scenario 1
MF(ACI)	6557	16.7119	–	–	–
ACI(ACI)	20,841	17.6964	14,284	0.9845	14,509
Scenario 2
MF(MF)	4974	16.6777	–	–	–
ACI(MF)	19,820	17.6182	14,845	0.9405	15,785

Figure 28a–j presents the CEACs for the SA for the different time horizons for scenarios 1 and 2. For the 10-year time horizon, scenario 1: if the decision-maker was willing to pay £20,000 then the probability that MF was more cost-effective than ACI was 62% (see Figure 28a); for scenario 2, MF was 64% more likely to be cost-effective than ACI (see Figure 28b). ACI became more cost-effective than MF when the decision-maker was willing to pay approximately £26,000 for scenario 1 and £28,000 for scenario 2. For all other time horizons, the probability that ACI was more cost-effective than MF was approximately 55% for both scenarios when the decision-maker was willing to pay £20,000. The results highlighted that, for the longer time horizons, ACI as a first repair was more cost-effective than MF as a first repair.

FIGURE 28.

Cost-effectiveness acceptability curves: different time horizons: (a) 10-year time horizon – scenario 1; (b) 10-year time horizon – scenario 2; (c) 20-year time horizon – scenario 1; (d) 20-year time horizon – scenario 2; (e) 30-year time horizon – scenario 1; (f) 30-year time horizon – scenario 2; (g) 40-year time horizon – scenario 1; (h) 40-year time horizon – scenario 2; (i) 50-year time horizon – scenario 1; and (j) 50-year time horizon – scenario 2.

Table 29 shows the deterministic and probabilistic cost-effectiveness results for the different time horizons and results were ranked by the least costly option. Again, the option MF(ACI) was extendedly dominated by a linear combination of MF(MF) and ACI(MF) and therefore this option was eliminated from the comparison. When comparing ACI(MF) with MF(MF) the deterministic ICER for a 10-year time horizon was over £27,000. For the same time period the deterministic ICER for the two initial ACI options: ACI(ACI) vs. ACI(MF) was approximately double this at £57,000. For the two initial ACI options, the deterministic ICER falls to just under £25,000 for a 20-year time horizon, and for the 30-year time horizon the ICER is just under £19,000. For both the 40- and 50-year time horizons (for the two initial ACI options), the deterministic ICER is very similar to the base-case ICER. The clear reason why the shorter time horizons are not cost-effective is due to the costs of ACI occurring at the start of the model and the benefits appearing much later, especially in terms of the reduced need for TKRs and fewer people going to the ‘no further repair’ or ‘no further knee replacement’ health states (where the utility is lower).

TABLE 29 - Sensitivity cost-effectiveness results: changing the time horizon

As the incremental QALYs are near zero, the ICER can fluctuate widely.

Procedure	Total mean costs (£)	Total mean QALYs	Comparison	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic: 10-year time horizon
MF(MF)	4498	7.2906	–	–	–	–
MF(ACI)	5983	7.3030	MF(ACI) vs. MF(MF)	1485	0.0124a	Extendedly dominated
ACI(MF)	19,329	7.8321	ACI(MF) vs. MF(MF)	14,831	0.5415	27,388
ACI(ACI)	20,082	7.8454	ACI(ACI) vs. ACI(MF)	753	0.0132a	56,816
Probabilistic: 10-year time horizon
MF(MF)	4505	7.2845	–	–	–	–
MF(ACI)	5989	7.2950	MF(ACI) vs. MF(MF)	1484	0.0105a	Extendedly dominated
ACI(MF)	19,326	7.8270	ACI(MF) vs. MF(MF)	14,821	0.5425	27,320
ACI(ACI)	20,075	7.8501	ACI(ACI) vs. ACI(MF)	749	0.0231a	32,448
Deterministic: 20-year time horizon
MF(MF)	4524	11.2587	–	–	–	–
MF(ACI)	6104	11.2812	MF(ACI) vs. MF(MF)	1580	0.0225a	Extendedly dominated
ACI(MF)	19,384	12.0654	ACI(MF) vs. MF(MF)	14,860	0.8067	18,421
ACI(ACI)	20,340	12.1040	ACI(ACI) vs. ACI(MF)	955	0.0386a	24,742
Probabilistic: 20-year time horizon
MF(MF)	4526	11.2362	–	–	–	–
MF(ACI)	6098	11.2630	MF(ACI) vs. MF(MF)	1572	0.0268a	Extendedly dominated
ACI(MF)	19,414	12.0625	ACI(MF) vs. MF(MF)	14,888	0.8263	18,019
ACI(ACI)	20,359	12.1136	ACI(ACI) vs. ACI(MF)	944	0.0511a	18,486
Deterministic: 30-year time horizon
MF(MF)	4739	13.9750	–	–	–	–
MF(ACI)	6329	13.9997	MF(ACI) vs. MF(MF)	1590	0.0247a	Extendedly dominated
ACI(MF)	19,609	14.8774	ACI(MF) vs. MF(MF)	14,871	0.9024	16,480
ACI(ACI)	20,614	14.9318	ACI(ACI) vs. ACI(MF)	1005	0.0544a	18,472
Probabilistic: 30-year time horizon
MF(MF)	4728	13.9748	–	–	–	–
MF(ACI)	6326	14.0117	MF(ACI) vs. MF(MF)	1598	0.0369a	Extendedly dominated
ACI(MF)	19,628	14.8754	ACI(MF) vs. MF(MF)	14,900	0.9006	16,545
ACI(ACI)	20,642	14.9494	ACI(ACI) vs. ACI(MF)	1014	0.0740a	13,705
Deterministic: 40-year time horizon
MF(MF)	4901	15.7354	–	–	–	–
MF(ACI)	6492	15.7604	MF(ACI) vs. MF(MF)	1591	0.0250a	Extendedly dominated
ACI(MF)	19,775	16.6747	ACI(MF) vs. MF(MF)	14,875	0.9393	15,836
ACI(ACI)	20,798	16.7368	ACI(ACI) vs. ACI(MF)	1022	0.0621a	16,466
Probabilistic: 40-year time horizon
MF(MF)	4900	15.7279	–	–	–	–
MF(ACI)	6494	15.7558	MF(ACI) vs. MF(MF)	1594	0.0276a	Extendedly dominated
ACI(MF)	19,736	16.6504	ACI(MF) vs. MF(MF)	14,837	0.9225	16,083
ACI(ACI)	20,763	16.7219	ACI(ACI) vs. ACI(MF)	1026	0.0716a	14,336
Deterministic: 50-year time horizon
MF(MF)	4987	16.6913	–	–	–	–
MF(ACI)	6579	16.7164	MF(ACI) vs. MF(MF)	1592	0.0251a	Extendedly dominated
ACI(MF)	19,864	17.6427	ACI(MF) vs. MF(MF)	14,876	0.9514	15,636
ACI(ACI)	20,891	17.7078	ACI(ACI) vs. ACI(MF)	1028	0.0651a	15,793
Probabilistic: 50-year time horizon
MF(MF)	4974	16.6777	–	–	–	–
MF(ACI)	6557	16.7119	MF(ACI) vs. MF(MF)	1583	0.0341a	Extendedly dominated
ACI(MF)	19,820	17.6182	ACI(MF) vs. MF(MF)	14,845	0.9405	15,785
ACI(ACI)	20,841	17.6964	ACI(ACI) vs. ACI(MF)	1022	0.0782a	13,073

Figure 29a–e presents the CEACs for the SA results for the different time horizons. The graphs suggest that for the 10-year time horizon, MF(MF) is the most cost-effective option if the decision-maker is willing to pay £30,000 per QALY. At over £36,000 per QALY, the most cost-effective option is ACI as a first repair – if a second repair is needed then this could either be ACI or MF. For the 20-year time horizon, MF(MF) appears to be the most cost-effective option if the decision-maker is willing to pay £22,000 per QALY. At over £26,000 per QALY, the most cost-effective option is ACI(ACI) – ACI as a first repair, and, if a second repair is needed, this should also be ACI. As the time horizon increases ACI(ACI) – ACI as a first repair, and, if a second repair is needed, this should also be an ACI – is probably more cost-effective than the other three sequences. So, for example, if the decision-maker is willing to pay £24,000 per QALY then the probability that ACI(ACI) is more cost-effective for both the 30- and 40-year time horizons is 28%, and the probability that ACI(ACI) is more cost-effective for the 50-year time horizon is 29%.

FIGURE 29.

Cost-effectiveness acceptability curve: changing the time horizon. (a) 10-year time horizon; (b) 20-year time horizon; (c) 30-year time horizon; (d) 40-year time horizon; and (e) 50-year time horizon.

Sensitivity analysis 3: microfracture done as a day case procedure

In the base-case analysis, according to clinical advice we have used a cost for MF as an inpatient procedure (£3020); however, we know that sometimes this procedure is done as day case. In the SA we have assumed that MF is done as a day case procedure at a cost of £1034. Table 30 shows the sensitivity cost-effectiveness results for MF as a day case procedure. The costs for MF have fallen, but the QALY gain does not change. Hence, ACI as a first repair is still the most cost-effective procedure compared with MF as a first repair, with ICERs of just over £16,000 (scenario 1) and just under £18,000 (scenario 2).

TABLE 30 - Sensitivity cost-effectiveness results (by scenario): MF procedure as a day case surgery and not as an inpatient

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic
Scenario 1
MF(ACI)	4621	17.0284	–	–	–
ACI(ACI)	20,921	18.0228	16,300	0.9944	16,391
Scenario 2
MF(MF)	2819	17.0033	–	–	–
ACI(MF)	19,756	17.9570	16,937	0.9537	17,758
Probabilistic
Scenario 1
MF(ACI)	4620	17.0412	–	–	–
ACI(ACI)	20,951	17.9975	16,332	0.9563	17,078
Scenario 2
MF(MF)	2811	17.0137	–	–	–
ACI(MF)	19,788	17.9065	16,977	0.8928	19,017

Figure 30a and b presents the CEACs for MF as a day case procedure for scenarios 1 and 2, respectively. If the decision-maker was willing to pay £20,000, for scenario 1 the probability that ACI was more cost-effective than MF was 55% (see Figure 30a), and for scenario 2 the probability that ACI was more cost-effective than MF was 52% (see Figure 30b).

FIGURE 30.

Cost-effectiveness acceptability curves: MF as a day case. (a) MF as a day case – scenario 1; and (b) MF as a day case – scenario 2.

Table 31 presents the deterministic and probabilistic cost-effectiveness results. Compared with the base-case analysis, even though the costs for MF have fallen, there is no change in the QALYs.

TABLE 31 - Sensitivity cost-effectiveness results: MF procedure as a day case surgery and not as an inpatient

As the incremental QALYs are near zero, the ICER can fluctuate widely.

Procedure	Total mean costs (£)	Total mean QALYs	Comparison	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic
MF(MF)	2819	17.0033	–	–	–	–
MF(ACI)	4621	17.0284	MF(ACI) vs. MF(MF)	1802	0.0251a	Extendedly dominated
ACI(MF)	19,756	17.9570	ACI(MF) vs. MF(MF)	16,937	0.9537	17,758
ACI(ACI)	20,921	18.0228	ACI(ACI) vs. ACI(MF)	1165	0.0658a	17,715
Probabilistic
MF(MF)	2811	17.0137	–	–	–	–
MF(ACI)	4620	17.0412	MF(ACI) vs. MF(MF)	1809	0.0275a	Extendedly dominated
ACI(MF)	19,788	17.9065	ACI(MF) vs. MF(MF)	16,977	0.8928	19,017
ACI(ACI)	20,951	17.9975	ACI(ACI) vs. ACI(MF)	1163	0.0910a	12,777

The ICERs between the different options were in line with the base-case results and initial ACI appears more cost-effective than initial MF, and, for those who need a second repair after the first ACI, this should be another ACI.

Figure 31 presents the CEAC and the graph highlights that if the decision-maker is not willing to pay more than £22,000 then MF(MF) is the most cost-effective option. If willing to pay more than £24,000, then ACI(MF) is the most cost-effective option – that is, the first repair should be ACI and if a second repair is needed this should be MF because of the lower costs (even though having an ACI as a second repair generates more QALYs).

FIGURE 31.

Cost-effectiveness acceptability curve: MF as a day case procedure.

Sensitivity analysis 4: improving the success rates of microfracture

In this SA we have conducted a ‘what if’ scenario, for which we have assumed that the duration of success for MF increases (1) by 20% and (2) by 40%. Table 32 shows the sensitivity cost-effectiveness results by scenario for the increase in the duration of success for MF. ACI was still more costly than MF and there were more QALYs gained with ACI than MF; even though there was a slight fall in the incremental QALYs gained compared with the base-case results, this was not enough to change the ICERs. For both scenarios, ACI as a first repair was more cost-effective than MF as a first repair. These results were of similar magnitudes and directions for the probabilistic results.

TABLE 32 - Sensitivity cost-effectiveness results (by scenario): improving the success rates of MF

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic: 20% increase in success rates
Scenario 1
MF(ACI)	6392	17.0756	–	–	–
ACI(ACI)	20,921	18.0228	14,529	0.9472	15,338
Scenario 2
MF(MF)	4969	17.0607	–	–	–
ACI(MF)	19,892	17.9639	14,923	0.9033	16,521
Probabilistic: 20% increase in success rates
Scenario 1
MF(ACI)	6387	17.0546	–	–	–
ACI(ACI)	20,895	18.0271	14,509	0.9725	14,919
Scenario 2
MF(MF)	4954	17.0403	–	–	–
ACI(MF)	19,856	17.9429	14,901	0.9027	16,508
Deterministic: 40% increase in success rates
Scenario 1
MF(ACI)	5698	17.0787	–	–	–
ACI(ACI)	20,921	18.0228	15,223	0.9441	16,125
Scenario 2
MF(MF)	4820	17.0758	–	–	–
ACI(MF)	19,892	17.9741	15,072	0.8983	16,778
Probabilistic: 40% increase in success rates
Scenario 1
MF(ACI)	5682	17.0535	–	–	–
ACI(ACI)	20,975	18.0331	15,292	0.9796	15,611
Scenario 2
MF(MF)	4803	17.0520	–	–	–
ACI(MF)	19,939	17.9887	15,136	0.9367	16,159

Figure 32a–d presents the CEACs for the SA for the increase in the duration of success for MF for scenarios 1 and 2. For a 20% increase in the duration of success for MF: scenario 1, if the decision-maker was willing to pay £20,000, the probability that ACI was more cost-effective than MF was 58% (see Figure 32a); scenario 2, the probability that ACI was more cost-effective than MF was 56% (see Figure 32b). For a 40% increase in the duration of success for MF these probability figures were 57% (see Figure 32c) and 55% (see Figure 32d), respectively.

FIGURE 32.

Cost-effectiveness acceptability curves: increase in MF success rate: (a) 20% increase in MF success rate – scenario 1; (b) 20% increase in MF success rate – scenario 2; (c) 40% increase in MF success rate – scenario 1; and (d) 40% increase in MF success rate – scenario 2.

Table 33 presents the deterministic and probabilistic cost-effectiveness results. The costs of MF have fallen slightly and also the QALYs for MF have increased. ACI(ACI) has an ICER of just under £18,000 (for a 20% increase in the duration of success rate for MF) and over £21,000 (for a 40% increase in the duration of success rate for MF).

TABLE 33 - Sensitivity cost-effectiveness results: improving the success rates of MF

As the incremental QALYs are near zero, the ICER can fluctuate widely.

Procedure	Total mean costs (£)	Total mean QALYs	Comparison	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic: 20% increase in success rates
MF(MF)	4969	17.0607	–	–	–	–
MF(ACI)	6392	17.0756	MF(ACI) vs. MF(MF)	1423	0.0149a	Extendedly dominated
ACI(MF)	19,892	17.9639	ACI(MF) vs. MF(MF)	14,923	0.9033	16,521
ACI(ACI)	20,921	18.0228	ACI(ACI) vs. ACI(MF)	1029	0.0589a	17,480
Probabilistic: 20% increase in success rates
MF(MF)	4954	17.0403	–	–	–	–
MF(ACI)	6387	17.0546	MF(ACI) vs. MF(MF)	1433	0.0144a	Extendedly dominated
ACI(MF)	19,856	17.9429	ACI(MF) vs. MF(MF)	14,901	0.9027	16,508
ACI(ACI)	20,895	18.0271	ACI(ACI) vs. ACI(MF)	1040	0.0842a	12,356
Deterministic: 40% increase in success rates
MF(MF)	4820	17.0758	–	–	–	–
MF(ACI)	5698	17.0787	MF(ACI) vs. MF(MF)	878	0.0029a	Extendedly dominated
ACI(MF)	19,892	17.9741	ACI(MF) vs. MF(MF)	15,072	0.8983	16,778
ACI(ACI)	20,921	18.0228	ACI(ACI) vs. ACI(MF)	1029	0.0487a	21,130
Probabilistic: 40% increase in success rates
MF(MF)	4803	17.0520	–	–	–	–
MF(ACI)	5682	17.0535	MF(ACI) vs. MF(MF)	880	0.0015a	Extendedly dominated
ACI(MF)	19,939	17.9887	ACI(MF) vs. MF(MF)	15,136	0.9367	16,159
ACI(ACI)	20,975	18.0331	ACI(ACI) vs. ACI(MF)	1036	0.0443a	23,356

Figure 33a and b presents the CEACs for the SA results for the ‘what if’ scenario if the duration of success of MF was to increase by 20% and 40%, respectively. For a 20% increase in the MF success rate, the graph suggests that if the decision-maker is willing to pay only £20,000 then there is not much difference in the four options; however, if the decision-maker is willing to pay £22,000 or more then ACI as a first repair is more cost-effective than MF as a first repair. For a 40% increase in the MF success rate, the graph indicates that if the decision-maker is willing to pay £20,000 then ACI as a first repair [ACI(ACI) or ACI(MF)] is more cost-effective than MF (approximately 32–33% probability that it is more cost-effective).

FIGURE 33.

Cost-effectiveness acceptability curves: increase in MF success rate. (a) Increase in MF success rate by 20%; and (b) increase in MF success rate by 40%. Note that MF only becomes cost-effective if the duration of benefit is much longer.

Sensitivity analysis 5: starting age of cohort is 45 years

In the base-case analysis the starting age for the cohort was 33 years. In this SA the starting age is changed to 45 years (patients are nearer to the knee replacement age) to see how this affects the ICER. Table 34 presents the deterministic and probabilistic cost-effectiveness results by scenario. Even though the model is starting at a later age (45 years), the results are very similar to the base case, with ACI as a first repair being cost-effective compared with MF as a first repair.

TABLE 34 - Sensitivity cost-effectiveness results (by scenario): starting age for cohort is 45 years

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic
Scenario 1
MF(ACI)	6422	15.0445	–	–	–
ACI(ACI)	16,784	16.0327	10,362	0.9882	10,486
Scenario 2
MF(MF)	5267	15.0187	–	–	–
ACI(MF)	16,116	15.9766	10,849	0.9579	11,326
Probabilistic
Scenario 1
MF(ACI)	6441	15.0833	–	–	–
ACI(ACI)	16,724	16.0377	10,283	0.9544	10,755
Scenario 2
MF(MF)	5281	14.9900	–	–	–
ACI(MF)	16,053	15.9562	10,772	0.9962	11,149

Figure 34a and b presents the CEACs for the SA with a starting age of 45 years for the cohort for scenarios 1 and 2. For scenarios 1 and 2, if the decision-maker was willing to pay £20,000, the probability that ACI was cost-effective relative to MF was 64%.

FIGURE 34.

Cost-effectiveness acceptability curves: starting age is 45 years: (a) scenario 1; and (b) scenario 2.

Table 35 shows the deterministic and probabilistic cost-effectiveness results with a starting age of 45 years for the cohort and results were ranked by the least costly option. Results were similar to the base-case results and ACI(ACI) remained the most cost-effective procedure.

TABLE 35 - Sensitivity cost-effectiveness results: starting age for cohort is 45 years

As the incremental QALYs are near zero, the ICER can fluctuate widely.

Procedure	Total mean costs (£)	Total mean QALYs	Comparison	Incremental costs (£)	Incremental QALYs	ICER (£) (cost per QALY gained)
Deterministic
MF(MF)	5267	15.0187	–	–	–	–
MF(ACI)	6422	15.0445	MF(ACI) vs. MF(MF)	1155	0.0258a	Extendedly dominated
ACI(MF)	16,116	15.9766	ACI(MF) vs. MF(MF)	10,849	0.9579	11,326
ACI(ACI)	16,784	16.0327	ACI(ACI) vs. ACI(MF)	667	0.0561a	11,898
Probabilistic
MF(MF)	5281	14.9900	–	–	–	–
MF(ACI)	6441	15.0833	MF(ACI) vs. MF(MF)	1160	0.0933a	Extendedly dominated
ACI(MF)	16,053	15.9562	ACI(MF) vs. MF(MF)	10,772	0.9662	11,149
ACI(ACI)	16,724	16.0377	ACI(ACI) vs. ACI(MF)	671	0.0815a	8241

Figure 35 presents the CEAC for a starting age of 45 years for the cohort. If the decision-maker is willing to pay £20,000 per QALY then either ACI(ACI) or ACI(MF) are the most cost-effective options.

FIGURE 35.

Cost-effectiveness acceptability curve: starting age of cohort is 45 years.

Methods and summary of findings

For the base-case analysis, for the discounted deterministic results MF was the less costly option but had fewer QALYs, whereas ACI was the more expensive option but generated more QALYs.

For scenario 1, the cost per QALY gained for ACI compared with MF was £14,395, and for scenario 2 the cost per QALY gained £15,598. These results were confirmed by the CEACs: so if the decision-maker is willing to pay £20,000 for a QALY, ACI is 56–59% more likely to be cost-effective than MF. For both scenarios, ACI as a first repair appeared more cost-effective than MF as a first repair.

When looking at the different sequences (options), the initial ACI appears more cost-effective than initial MF, and, for those who need a second repair after the first ACI, this should also be another ACI.

The option MF(ACI) was extendedly dominated by a linear combination of MF(MF) and ACI(MF) and therefore this option was eliminated from the comparison. For the different sequences, the CEAC for the base-case results confirmed these results and showed that if the decision-maker is willing to pay £18,000 or more for a QALY than ACI as a first procedure is more cost-effective than MF as a first procedure.

We found that the key cost driver was the cost of the cells for the ACI procedure, but, over the time horizon, ACI is more beneficial (more gain in QALYs) and cost saving to the NHS (fewer people in need of a second repair or of a TKR).

A number of SAs were undertaken to determine the cost-effectiveness of various options and the majority of results were in line with the base-case analysis. However, we found that the model was sensitive to the cost of cells – we know that these are not the true costs as the NHS receives confidential discounts from the manufacturers. This means that with the cell cost reduction ACI(ACI) is likely to be even more cost-effective than the base-case cost per QALY ratio that was presented. We also found that the model was sensitive to the time horizon: with a shorter time horizon – 10 years – the cost per QALY for the two initial ACI options rose to around £26,000 because of the costs of the ACI procedure occurring at the start and the benefits of ACI not being realised until much later, such as the reduced need for TKRs. When the time horizon was longer, the model results were in line with the base-case results. The SAs conducted using Oswestry data found ACI not to be cost-effective compared with MF and mainly due to the lower utility value in the fourth year for ACI compared with MF. However, for reasons explained in Chapter 5, there are several confounding factors that influence these utility values.

Strengths and limitations

If the first repair fails, the Markov model considers patients having a maximum of two knee repairs (any combination of MF and ACI) if they choose to.

However, the model does have a number of limitations. First, the length of follow-up we found in the trials published in the literature was too short and hence, there are no long-term data on the success and failure rates (including long-term benefits and AEs) for each of these procedures and what the average age is for these patients when a TKR/PKR is required. Hence, we relied heavily on two papers70,100 to obtain transition probabilities for our economic model. In addition, because of the paucity of data from clinical studies, transition probabilities were not available for each transition in the model; hence we made a number of assumptions for our model, which were checked for plausibility with our clinical advisors. However, results from the long-term ACTIVE trial35 (comparing ACI/MACI with standard treatments) and the TOPKAT trial (comparing TKR with PKR)164 will provide useful information with which to populate our economic model, although results will not be available until 2017 and 2019, respectively.

Second, because of the short follow-up, we also found that there were no long-term data on utility values that were associated with each of these procedures. We have had to rely heavily on the literature and on a few studies in particular, such as Gerlier et al. 141 Also, we found no studies that mapped any of the clinical measures – such as Lysholm score or the KOOS – to either the EQ-5D or SF-6D to generate utility values, which would have been helpful in our model.

Third, we relied on clinical experience for information on the average resources used (e.g. outpatient and rehabilitation visits) for these patients.

Fourth, we did not take into account any costs for the analgesics, based on advice from our clinical experts, as these costs are negligible and would not have altered the base-case cost per QALY. Also, not all of the costs obtained were from the NHS reference costs. We used the previous HTA report by Clar et al. ,3 who obtained costs from Aberdeen and Southampton hospitals to populate their economic model. Although these costs were inflated to 2012–13 prices using the HCHS index,154 to get a more accurate picture of these costs it would have been better to have carried out ‘bottom-up costing’.

Fifth, the model has not taken into account any private patient costs, such as time off work and loss of pay (productivity): this population, of those who have either an ACI or MF, is a young cohort and it will primarily have an effect on their own costs. In line with this, it would have been interesting to know how long it would take this population cohort to return to normal activities after each of these procedures (return to work or sports).

Finally, we did not include any AEs, as there were no key differences between the two treatment arms.

New base case

• Data used for ACI failure rates: Nawaz et al. 80 (whole cohort).

• Data used for MF failure rates: pooled data (Layton 2015,114 Knutsen 2007,67 Saris 200970).

• Cost of harvesting: £870.

• Cost of implantation: £2396, as requested by NICE. This includes an inpatient stay. ACI can be done on a day case basis, though it should be noted that, because it is often provided as a specialist ‘regional’ service, overnight stays may be unavoidable because of distance. The clinical authors of this report vary between one-night stays for all and some being done as day case. The operation is often open, and such exposure is much more painful than the arthroscopic surgery used for harvesting the initial tissue). However, mini-arthrotomy may be used.

• MF is nearly always a day case procedure.

Table 36 presents the results from the new base-case analysis. The discounted deterministic results show that MF(ACI) was the least costly option and had the fewest QALYs; although ACI(ACI) generated the most QALYs, it was also the most expensive option. The option of ACI(MF) was extendedly dominated by a linear combination of MF(ACI) and ACI(ACI), and therefore this option was eliminated from the comparison. The ICER comparing ACI(ACI) with MF(ACI) was just under £19,000; doing ACI first is more cost-effective. The discounted probabilistic results were very similar. Figure 36 shows the cost-effectiveness analysis for the two remaining options. The graph shows that, for amounts below £20,000, MF(ACI) is the most cost-effective option; at a willingness to pay of £20,000 there is not much difference between the two options, and, at a willingness to pay above £20,000, ACI(ACI) is probably more cost-effective.

FIGURE 36.

Cost-effectiveness acceptability curve: base case.

Sensitivity analyses on prices

Note that in analysis of price changes, the QALY gain does not change in the deterministic arms, as expected. However, when the model is run probabilistically all of the input variables change because of the different distributions, hence both the costs and QALYs will change.

Cells at £6000

• Data used for ACI failure rates: Nawaz et al. 80 (whole cohort).

• Data used for MF failure rates: pooled data (Layton 2015,114 Knutsen 2007,67 Saris 200970).

• Cost of harvesting: £870.

• Cost of implantation: £2396.

• Cost of cells: £6000.

Table 37 shows the results when the price of cells is reduced to £6000. The deterministic discounted results highlighted that the option of ACI(MF) was extendedly dominated and therefore the ICER comparing ACI(ACI) with MF(ACI) was just under £7500. Figure 37 shows that at a willingness to pay of £8000 there is not much difference between the two remaining options.

TABLE 37 - Deterministic and probabilistic cost-effectiveness results: £6000 price

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER vs. MF(ACI)	Comparator
Deterministic: undiscounted
MF(ACI)	6771	34.2885	–	–	–	–
ACI(MF)	12,661	35.5596	5,890	1.2711	Extended dominated	MF(ACI)
ACI(ACI)	13,244	35.6999	583	0.1403	4586	MF(ACI)
Deterministic: discounted
MF(ACI)	5441	17.1350	–	–	–	–
ACI(MF)	11,400	17.9304	5,959	0.7954	Extended dominated	MF(ACI)
ACI(ACI)	11,820	17.9953	420	0.0650	7414	MF(ACI)
Probabilistic: discounted
MF(ACI)	5452	17.1340	–	–	–	–
ACI(MF)	11,486	17.9110	6,034	0.7770	7766	MF(ACI)
ACI(ACI)	11,909	17.9474	423	0.0364	11,622	ACI(MF)

FIGURE 37.

Cost-effectiveness acceptability curve: £6000.

Cells at cost £8000

• Data used for ACI failure rates: Nawaz et al. 80 (whole cohort).

• Data used for MF failure rates: pooled data (Layton 2015,114 Knutsen 2007,67 Saris 200970).

• Cost of harvesting: £870.

• Cost of implantation: £2396.

• Cost of cells: £8000.

Table 38 shows the results when the price of cells is reduced to £8000. The deterministic discounted results highlighted that the option of ACI(MF) was extendedly dominated and therefore the ICER comparing ACI(ACI) with MF(ACI) was just under £10,000. Figure 38 shows that at a willingness to pay of £10,000 there is not much difference between the two remaining options.

TABLE 38 - Deterministic and probabilistic cost-effectiveness results: £8000 price

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER vs. MF(ACI)	Comparator
Deterministic: undiscounted
MF(ACI)	6964	34.2885	–	–	–	–
ACI(MF)	14,661	35.5596	7697	1.2711	Extended dominated	MF(ACI)
ACI(ACI)	15,422	35.6999	761	0.1403	5993	MF(ACI)
Deterministic: discounted
MF(ACI)	5602	17.1350	–	–	–	–
ACI(MF)	13,400	17.9304	7797	0.7954	Extended dominated	MF(ACI)
ACI(ACI)	13,948	17.9953	549	0.0650	9700	MF(ACI)
Probabilistic: discounted
MF(ACI)	5608	17.1630	–	–	–	–
ACI(MF)	13,430	17.9500	7822	0.7871	9938	MF(ACI)
ACI(ACI)	13,983	18.0242	553	0.0742	7454	ACI(MF)

FIGURE 38.

Cost-effectiveness acceptability curve: £8000.

Cells at cost of £12,000

• Data used for ACI failure rates: Nawaz et al. 80 (whole cohort).

• Data used for MF failure rates: pooled data (Layton 2015,114 Knutsen 2007,67 Saris 200970).

• Cost of harvesting: £870.

• Cost of implantation: £2396.

• Cost of cells: £12,000.

Table 39 shows the results when the price of cells is reduced to £12,000. The deterministic discounted results highlighted that the option of ACI(MF) was extendedly dominated and therefore the ICER comparing ACI(ACI) with MF(ACI) was just under £14,500. Figure 39 shows that at a willingness to pay of £16,000 there is not much difference between the two remaining options.

TABLE 39 - Deterministic and probabilistic cost-effectiveness results: £12,000 price

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER vs. MF(ACI)	Comparator
Deterministic: undiscounted
MF(ACI)	7350	34.2885	–	–	–	–
ACI(MF)	18,661	35.5596	11,312	1.2711	Extended dominated	MF(ACI)
ACI(ACI)	19,778	35.6999	1117	0.1403	8806	MF(ACI)
Deterministic: discounted
MF(ACI)	5925	17.1350	–	–	–	–
ACI(MF)	17,400	17.9304	11,475	0.7954	Extended dominated	MF(ACI)
ACI(ACI)	18,205	17.9953	805	0.0650	14,272	MF(ACI)
Probabilistic: discounted
MF(ACI)	5918	17.1425	–	–	–	–
ACI(MF)	17,320	17.9539	11,402	0.8114	Extended dominated	MF(ACI)
ACI(ACI)	18,131	17.9899	811	0.0360	14,412	MF(ACI)

FIGURE 39.

Cost-effectiveness acceptability curve: £12,000.

Sensitivity analyses: post-repair utility

In our first assessment report, we assumed that patients who decided not to have a further repair had had some benefit, and had improved from a utility of 0.654 before the repair to 0.691 afterwards. NICE asked us to assess the effect of several assumptions for utilities in those in whom repair is unsuccessful, but who choose not to have another operation.

Utility for choose no second repair set to same as failure

• Data used for ACI failure rates: Nawaz et al. 80 (whole cohort).

• Data used for MF failure rates: pooled data (Layton 2015,114 Knutsen 2007,67 Saris 200970).

• Cost of harvesting: £870.

• Cost of implantation: £2396.

• Cost of cells: £16,000.

Utility for those who choose no second repair: 0.654.

Table 40 shows the results when the utility value for those who choose no second repair is 0.654. The deterministic discounted results highlighted that the option of ACI(MF) was extendedly dominated and therefore the ICER comparing ACI(ACI) with MF(ACI) was just over £15,500 (results are similar to the new base-case results). Figure 40 shows that at a willingness to pay of £16,000 there is not much difference between the two remaining options.

TABLE 40 - Deterministic and probabilistic cost-effectiveness results: utility = 0.654

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER vs. MF(ACI)	Comparator
Deterministic: undiscounted
MF(ACI)	7736	32.8665	–	–	–	–
ACI(MF)	22,661	34.4351	14,926	1.5686	Extended dominated	MF(ACI)
ACI(ACI)	24,134	34.6021	1473	0.1670	9449	MF(ACI)
Deterministic: discounted
MF(ACI)	6248	16.5058	–	–	–	–
ACI(MF)	21,400	17.4667	15,152	0.9609	Extended dominated	MF(ACI)
ACI(ACI)	22,461	17.5428	1062	0.0762	15,634	MF(ACI)
Probabilistic: discounted
MF(ACI)	6247	16.4820	–	–	–	–
ACI(MF)	21,446	17.4583	15,198	0.9763	Extended dominated	MF(ACI)
ACI(ACI)	22,2522	17.5499	1077	0.0916	15,241	MF(ACI)

FIGURE 40.

Cost-effectiveness acceptability curve: utility = 0.654.

Utility for choose no second repair set to same as success

• Data used for ACI failure rates: Nawaz et al. 80 (whole cohort).

• Data used for MF failure rates: pooled data (Layton 2015,114 Knutsen 2007,67 Saris 200970).

• Cost of harvesting: £870.

• Cost of implantation: £2396.

• Cost of cells: £16,000.

• Utility for those who choose no second repair: 0.817. Note that this assumption greatly increases utility gain among those who do not get good results after MF, and reduces the marginal QALY gains from ACI.

Table 41 shows the results when the utility value for those who choose no second repair is 0.817. Here we are assuming that there are increases in utility gain among those who do not get good results after MF, and therefore this reduces the marginal QALY gains from ACI. The deterministic discounted results highlighted that the option of ACI(MF) was extendedly dominated and therefore the ICER comparing ACI(ACI) with MF(ACI) was just over £62,000. Hence, the option ACI(ACI) is no longer considered to be cost-effective. Figure 41 shows that, at a willingness to pay of £20,000, MF(ACI) is 70% more likely to be cost-effective than ACI(ACI).

TABLE 41 - Deterministic and probabilistic cost-effectiveness results: utility = 0.817

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER vs. MF(ACI)	Comparator
Deterministic: undiscounted
MF(ACI)	7736	39.1309	–	–	–	–
ACI(MF)	22,661	39.3889	14,926	0.2580	Extended dominated	MF(ACI)
ACI(ACI)	24,134	39.4383	1473	0.0494	53,352	MF(ACI)
Deterministic: discounted
MF(ACI)	6248	19.2776	–	–	–	–
ACI(MF)	21,400	19.5096	15,152	0.2320	Extended dominated	MF(ACI)
ACI(ACI)	22,461	19.5363	1062	0.0267	62,658	MF(ACI)
Probabilistic: discounted
MF(ACI)	6250	19.2966	–	–	–	–
ACI(MF)	21,438	19.5140	15,188	0.2290	Extended dominated	MF(ACI)
ACI(ACI)	22,512	19.5484	1074	0.0384	64,581	MF(ACI)

FIGURE 41.

Cost-effectiveness acceptability curve: utility = 0.817.

Utility for choose no second repair set to mid-point of success and failure

• Data used for ACI failure rates: Nawaz et al. 80 (whole cohort).

• Data used for MF failure rates: pooled data (Layton 2015,114 Knutsen 2007,67 Saris 200970).

• Cost of harvesting: £870.

• Cost of implantation: £2396.

• Cost of cells: £16,000.

Utility for those who choose no second repair: 0.746. This also reduces the marginal QALY gains from ACI as first procedure because the larger proportion who do not do well after MF have their utility increased.

Table 42 shows the results when the utility value for those who choose no second repair is 0.746. The deterministic discounted results highlighted that the option of ACI(MF) was extendedly dominated and therefore the ICER comparing ACI(ACI) with MF(ACI) was just over £27,000. Figure 42 shows that at a willingness to pay of £26,000 there is not much difference between the two remaining options.

TABLE 42 - Deterministic and probabilistic cost-effectiveness results: utility = 0.746

Procedure	Total mean costs (£)	Total mean QALYs	Incremental costs (£)	Incremental QALYs	ICER vs. MF(ACI)	Comparator
Deterministic: undiscounted
MF(ACI)	7736	36.4022	–	–	–	–
ACI(MF)	22,661	37.2311	14,926	0.8289	Extended dominated	MF(ACI)
ACI(ACI)	24,134	37.3317	1473	0.1006	17,643	MF(ACI)
Deterministic: discounted
MF(ACI)	6248	18.0702	–	–	–	–
ACI(MF)	21,400	18.6197	15,152	0.5495	Extended dominated	MF(ACI)
ACI(ACI)	22,461	18.6680	1062	0.0483	27,123	MF(ACI)
Probabilistic: discounted
MF(ACI)	6248	18.0400	–	–	–	–
ACI(MF)	21,419	18.6257	15,171	0.5857	Extended dominated	MF(ACI)
ACI(ACI)	22,496	18.6684	1077	0.0427	25,857	MF(ACI)

FIGURE 42.

Cost-effectiveness acceptability curve: utility = 0.746.

Subgroup analyses