Colonoscopy surveillance following adenoma removal to reduce the risk of colorectal cancer: a retrospective cohort study

Polyp numbering

We assigned a unique polyp number to polyps seen at more than one examination so that different sightings could be linked. We thought that different sightings were likely to be of the same polyp when sightings were in the same or adjacent colonic segment, there was an indication that a polyp seen at an earlier examination had not been completely removed, bowel preparation quality at an earlier examination was poor and/or the grades of dysplasia and histology reported at the different sightings were similar. We assigned a match probability to sightings to indicate our confidence that they were of the same polyp. Match probabilities were estimated to the nearest 10%. We considered sightings matched with a probability of ≥ 70% to be of the same polyp.

Polyp groups

When polyps were described as groups rather than as individual polyps we created a single record for each polyp group. We then populated the record with information on the number, size, shape, histopathology and location of polyps in the group. We assigned a unique group number to each polyp group and a group-linking number to polyp groups seen at more than one examination. When information was recorded for an individual polyp within a group we created an individual polyp record and linked it to the group record.

Terms such as ‘some’, ‘several’ and ‘many’ were often used to describe the number of polyps in a group. We assigned numeric values to these descriptive terms (e.g. some = 3), using the median value calculated from endoscopy reports that gave both a descriptive term and an exact polyp number. Once we had an estimate for the number of polyps in each group we created individual polyp records from the group records. Information recorded for the group was replicated in the individual polyp records.

Summary values for polyp characteristics

Polyp characteristics were often described in multiple data fields on the study database, particularly if the polyp was seen at more than one examination. In such cases, we combined all available data to create summary values for polyp size, shape, histology and location, using hierarchies of rules.

Colonic examinations

Outcomes and exposures

The primary outcome measures were long-term CRC incidence after baseline and the first surveillance visit and detection rates of AAs and CRC at the first surveillance visit. Outcomes of AA and CRC were ascertained using pathology data stored on the study database and cancer data from national data sources (for CRC).

We defined AAs as adenomas ≥ 10 mm in size or with villous or tubulovillous histology, or high-grade dysplasia. We defined CRC sites by the International Classification of Diseases revisions 8, 9 and 10,24–26 including codes C18–C20. We defined CRC morphology by the Manual of Tumor Nomenclature and Coding27 and the International Classification of Diseases for Oncology revisions 1 and 2. 28,29 We included adenocarcinomas of the colorectum as CRC outcomes, as well as cancers with unspecified morphology but assumed to be adenocarcinomas (i.e. those located between the rectum and caecum). Cancers with unspecified morphology but assumed to be squamous cell carcinomas (i.e. those located around the anus) and in situ cancers were not included as CRC outcomes.

We excluded CRCs that we assumed had developed from lesions that were incompletely resected at baseline. We did this because we thought that their inclusion could bias our estimates of CRC risk and lead to inappropriate surveillance recommendations. We assumed that CRCs had arisen from an incompletely resected lesion if all of the below criteria were met:

• The CRC was found in the same or adjacent colonic segment to a baseline lesion.

• The baseline lesion was an adenoma.

• The baseline adenoma was ≥ 15 mm in size.

• The baseline adenoma was seen on two or more occasions within 5 years before the cancer diagnosis.

In a sensitivity analysis we excluded some additional cancers that met some but not all the criteria above that we thought were likely to have arisen from an incompletely resected baseline lesion.

For our analyses of findings at the first surveillance visit we excluded all AAs seen during the baseline visit. We did this because we had previously showed that adenomas detected at first surveillance that were also seen at baseline were likely to be under polypectomy site surveillance and that the inclusion of such adenomas confounded analyses. 23

The primary exposures of interest were the number of surveillance visits and the length of the surveillance interval from baseline to the first surveillance visit. Patient follow-up was censored at first CRC diagnosis, first diagnosis of a condition affecting colonic surveillance regimen (including IBD, colitis, hyperplastic polyposis, proctitis or volvulus), performance of bowel resection, death, emigration, or the date when cancer registration data were considered complete (for patients matched to national data sources) or when the last examination was recorded on the study database (for patients not matched to national data sources).

In all analyses we excluded surveillance visits that fully occurred after censoring (i.e. if all examinations in the visit occurred after censoring). In our analyses of long-term CRC incidence, visits at which a censoring event occurred and visits with the last examination occurring after the date of censoring were not included as surveillance visits, as they did not offer protection against the development of CRC. By contrast, in our analyses of findings at first surveillance we included any first surveillance visits at which a censoring event occurred.

We defined the surveillance interval as the time period between the latest most complete examination in one visit to the first examination in the next visit. The surveillance interval was represented as a categorical variable and patients with the shortest surveillance interval were the reference group against which we compared patients exposed to a longer interval. In the case of the low- and intermediate-risk groups, the surveillance interval was categorised into the following: < 18 months, 2 years (± 6 months), 3 years (± 6 months), 4 years (± 6 months), 5 years (± 6 months), 6 years (± 6 months), 7 years (± 6 months), 8 years (± 6 months), 9 years (± 6 months) and ≥ 9.5 years. The following categories were used for the high-risk group: < 15 months, 1.5 years (± 3 months), 2 years (± 3 months), 2.5 years (± 3 months), 3 years (± 3 months), 3.5 years (± 3 months), 4 years (± 3 months), 4.5 years (± 3 months), 5 years (± 3 months) and ≥ 5.25 years.

Risk factors and confounders

We assessed the following patient, procedural and polyp characteristics as potential risk factors and confounders at baseline: age, sex, year of visit, length of visit (in days or months), examination completeness, bowel preparation quality, number of adenomas, adenoma size, histology, dysplasia, presence of proximal polyps, presence of hyperplastic polyps, presence of a large (i.e. ≥ 10 mm in size) hyperplastic polyp and family history of cancer/CRC. In a sensitivity analysis we additionally considered the hospital that patients attended as a confounding variable.

Family history of cancer/CRC was defined as ‘family history of cancer or CRC reported at an examination before or during visit’. Of cases reported to have a ‘family history of cancer’, 72% came from a specialist hospital for colorectal diseases and we therefore assumed that these cases had a family history of CRC.

We created categorical variables for the following continuous quantitative variables: number of surveillance visits, age, length of visit, number of adenomas and adenoma size. We created an unknown category for variables with missing data.

Long-term colorectal cancer incidence after baseline

We estimated the long-term incidence of CRC after baseline. Time at risk started from the last examination at baseline. Time-to-event data were censored at first CRC diagnosis; first diagnosis of IBD, colitis, hyperplastic polyposis, proctitis or volvulus; performance of bowel resection; death; emigration; or the date of complete case ascertainment in cancer registries. Patients who were not matched to national data sources were censored at the date of their last examination recorded on the study database rather than the date of complete case ascertainment.

We assessed the effects of baseline characteristics (see Risk factors and confounders) and surveillance on CRC incidence using Cox proportional-hazards models to calculate hazard ratios (HRs) and 95% confidence intervals (CIs). Exposure to successive surveillance visits started at the last examination in each visit.

We used multivariable Cox proportional hazards models to identify baseline CRC risk factors (i.e. characteristics independently associated with increased long-term CRC incidence). This involved using a backward stepwise selection procedure to retain variables with p-values of < 0.05 in the likelihood ratio test (LRT). We included the number of surveillance visits as a time-varying covariate, meaning that patients who had any surveillance contributed person-years to more than a single category of number of surveillance visits. We investigated interactions between the number of surveillance visits and age or sex by fitting the models with interaction parameters.

Higher- and lower-risk subgroups

We used the identified baseline CRC risk factors to divide each of the three risk groups into higher- and lower-risk subgroups. Patients with any of the baseline risk factors were assigned to the higher-risk subgroup, whereas those with none of the risk factors were assigned to the lower-risk subgroup. We did not include age in the risk classification criteria because older age is associated with poorer colonoscopy quality and greater risks of colonoscopy-related complications. 30 We also did not include the year or length of the baseline visit as these factors do not help define clinically meaningful patient subgroups.

We conducted a sensitivity analysis of the risk classification criteria for intermediate-risk patients. This involved using the baseline risk factors identified in our previous study of the intermediate-risk group in the classification of higher risk (i.e. incomplete colonoscopies or colonoscopies of unknown completeness, poor bowel preparation, adenomas ≥ 20 mm in size or with high-grade dysplasia, and proximal polyps). 20,23

Cumulative incidence of colorectal cancer

We used one minus the Kaplan–Meier estimator of the survival function to show time to CRC diagnosis and to estimate cumulative incidence of CRC with 95% CIs at 3, 5 and 10 years. We used the log-rank test to compare cumulative incidence curves.

Comparisons with colorectal cancer incidence in the general population

We compared CRC incidence with that in the general population by calculating standardised incidence ratios (SIRs). These were calculated as the ratio of observed to expected cases of CRC, with exact Poisson 95% CIs. We estimated the number of expected cases by multiplying the observed sex- and 5-year age group-specific number of person-years by the corresponding CRC incidence in the general population of England in 2007. 31

For our analyses of CRC incidence we divided each patient’s follow-up time into four distinct blocks of time:

1. without surveillance (i.e. from start of time-at-risk, censored at any first surveillance)

2. after first surveillance (i.e. from first surveillance, censored at any second surveillance)

3. after second surveillance (i.e. from second surveillance, censored at any third surveillance)

4. after third surveillance (i.e. from third surveillance, censored at end of follow-up).

In some analyses we combined the last two blocks of time to show CRC incidence in the presence of two or more surveillance visits. We compared CRC incidence in the presence of one surveillance visit with two or more surveillance visits using univariable Cox proportional hazards models.

Findings at the first surveillance visit

We assessed detection rates of AAs and CRC at the first surveillance visit by interval length (from baseline to first surveillance), overall and by risk subgroup. A test for trend was used to test for association between detection rates and interval length. Univariable logistic regression was used to calculate unadjusted odds ratios (ORs) and 95% CIs, comparing the detection of AAs and CRC between risk subgroups.

For the analyses of AA detection rates we excluded patients in whom CRC was detected at first surveillance because we expected that their risk of having an AA detected would be different from that of patients without CRC and that their inclusion could therefore lead to biased estimates.

Within-study analysis

We first conducted a within-study analysis using resource use and outcomes data from the main study. We considered resource use associated with both adenoma surveillance and treatment of CRC. Surveillance procedures included diagnostic and therapeutic colonoscopies, flexible sigmoidoscopies and rigid sigmoidoscopies. Unit costs for these procedures were taken from the NHS national schedule of reference costs for 2017–18. 33 We estimated the lifetime costs of CRC treatment using published estimates from a whole-disease model of CRC,34 inflating these costs from 2012/13 to 2017/18 prices using the gross domestic product deflator. Treatment costs varied according to patient age and CRC stage at diagnosis. We handled missing CRC staging data by means of multiple imputation using an ordered logit model. We discounted costs and the number of CRC cases at a rate of 3.5% per year.

For each of the three main risk groups we calculated the following for the higher- and lower-risk subgroups:

• incremental cost associated with adopting surveillance compared with no surveillance per 1000 person-years

• incremental number of CRCs diagnosed among patients attending surveillance compared with those not attending surveillance

• incremental cost per CRC prevented by adopting surveillance compared with no surveillance.

Lifetime analysis

For the lifetime analysis we used an extrapolation model to extrapolate results from the within-study analysis over a lifetime horizon. We designed a Markov model that consisted of eight states: (1) no surveillance visits, (2) surveillance visits, (3) Dukes’ stage A CRC, (4) Dukes’ stage B CRC, (5) Dukes’ stage C CRC, (6) Dukes’ stage D CRC, (7) death from CRC and (8) death from other causes. We estimated time-homogeneous probabilities of transitions between states for each risk group. We used multiple data sources to estimate these probabilities, including the main study database, the 2013 National Bowel Cancer Audit Annual Report35 and the Office for National Statistics. 36

We assigned quality-of-life (QoL) estimates to each state. For the CRC states we used mean EuroQol-5 Dimensions (EQ-5D) scores from a study37 that collected patient-reported outcome measures (PROMs) from patients with CRC. For the non-cancer states we used EQ-5D scores from a study38 that pooled responses in the Health Survey for England from people without cancer.

We ran the Markov model for the higher- and lower-risk subgroups of each risk group. This allowed us to calculate the costs and QALYs associated with surveillance. We discounted future costs and QALYs at an annual rate of 3.5%. We calculated incremental cost-effectiveness ratios (ICERs) as the ratio between the mean difference in QALYs and the mean difference in costs. Cost-effectiveness was evaluated assuming a willingness-to-pay threshold of £20,000 per QALY gained.

Previous approvals for our original study of the intermediate-risk group

We initially obtained research governance approvals to permit data collection from hospitals and national databases for our original study of the intermediate-risk group. These approvals are described in full in the NIHR final report for the original study23 and are summarised below:

• Ethics approval was granted by the Royal Free Research Ethics Committee (reference 06/Q0501/45).

• Approval for the processing of patient-identifiable information without consent in England was granted by the Patient Information Advisory Group (PIAG) under Section 60 of the Health and Social Care Act 200139 (re-enacted by Section 251 of the NHS Act 200640) (reference PIAG 1–05[e]/2006).

• Approval was granted by the Community Health Index Advisory Group of NSS to access the Community Health Index. This enabled the Information Services Division of the NSS to link patient-identifiable information with data from cancer and death registries.

• Research and development approval was obtained for all participating hospitals.

Approvals for the present study of all three risk groups

We subsequently obtained additional ethics approval from the London – Hampstead Research Ethics Committee and the Health Research Authority (reference 06/Q0501/45, IRAS ID 59943) for substantial amendments that extended the scope of the study protocol to examine the low- and high-risk groups, in addition to the intermediate-risk group previously analysed. Amendments and annual reviews of our approval to process patient-identifiable information without consent were approved by the Health Research Authority Confidentiality Advisory Group (reference PIAG 1–05[e]/2006).

Our data-sharing agreements with NHS Digital (reference DARS-NIC-147827-NC2TC) and NHS NSS (reference PBPP 1718–0048/SR244) were amended to include these additional analyses, and a new application was made to the Public Health England (PHE) Office for Data Release (reference ODR1718_326) to obtain cancer staging and treatment data for the health economic analysis.

We renewed our Data Access Agreements to allow continued access to data from the participating hospitals through December 2022.

To protect patient confidentiality, all information kept at the Cancer Screening and Prevention Group’s office was pseudo-anonymised and no patient-identifiable information, except for date of birth, was stored on the study database (Oracle database, Oracle Corporation, Redwood City, CA, USA). Access controls to the study database are in place, including password control and a firewall that limited access to a subset of Cancer Screening and Prevention Group computers.

Role of the funding source

The NIHR stipulated that we use a retrospective cohort study design, but it had no involvement in data collection, analysis or interpretation, in the writing of the report or in the decision to submit the report for publication.

The low-risk group

The median age of the low-risk group was 64 [interquartile range (IQR) 55–72] years, and 8019 (56%) patients were men. Approximately half of the group (n = 7194) were counted as having attended one or more surveillance visit in our analysis of long-term CRC incidence (Table 1). Patients who attended surveillance were younger and more likely to have had a baseline visit before 2005 or a baseline visit ≥ 6 months in length than those who did not attend surveillance. Non-attenders were more likely than attenders to have had an incomplete baseline colonoscopy and poor bowel preparation. Attenders were more likely to have had two adenomas (rather than one), an adenoma with tubulovillous or villous histology, an adenoma with high-grade dysplasia or hyperplastic polyps at baseline or to have a family history of cancer/CRC. The proportion of patients in whom data on adenoma histology, adenoma dysplasia, colonoscopy completeness and bowel preparation quality were missing was higher among attenders than among non-attenders (see Table 1).

The intermediate-risk group

The median age of the intermediate-risk group was 66 (IQR 58–74) years, and 6581 (56%) were men. A total of 7169 (60%) patients attended one or more surveillance visit during the long-term CRC incidence analysis (Table 2). Compared with patients who did not attend surveillance, attenders included a higher proportion of men and patients aged < 75 years, and were more likely to have had a baseline visit before 2005 and to have had a longer baseline visit (> 1 day). Non-attenders were more likely than attenders to have had an incomplete colonoscopy or poor bowel preparation at baseline. Attenders were more likely to have had an adenoma ≥ 20 mm in size, with tubulovillous histology, or with high-grade dysplasia, hyperplastic polyps, a hyperplastic polyp ≥ 10 mm in size at baseline or to have a family history of cancer/CRC. A greater proportion of attenders than of non-attenders were missing data for adenoma histology, adenoma dysplasia, colonoscopy completeness and bowel preparation quality (see Table 2).

The high-risk group

The median age of the high-risk group was 67 (IQR 61–74) years, and 1920 (71%) patients were men. In total, 1808 (66%) patients attended one or more surveillance visit during the long-term CRC incidence analysis (Table 3). Patients who attended surveillance were, on average, younger than non-attenders and more likely to have had a baseline visit before 2000 and a longer baseline visit (≥ 3 months). Non-attenders were more likely than attenders to have had an incomplete colonoscopy or poor bowel preparation at baseline. Attenders were more likely to have had six or more adenomas or hyperplastic polyps at baseline, and were more likely to have a family history of cancer/CRC. The proportion of patients for whom data for colonoscopy completeness and bowel preparation quality were missing was higher among attenders than among non-attenders (see Table 3).

The low-risk group

In the low-risk group, 195 CRCs were diagnosed during 138,903 person-years of follow-up (median 9.6 years, IQR 7.2–12.4 years), giving an incidence rate of 140 per 100,000 person-years (95% CI 122 to 162 per 100,000 person-years). In multivariable regression analysis, number of surveillance visits, age, completeness of colonoscopy, adenoma histology and proximal polyps were independently associated with CRC incidence. Adjusting for these factors, one surveillance visit was associated with a 44% reduction in CRC incidence compared with no surveillance (HR 0.56, 95% CI 0.39 to 0.80). Even greater reductions in incidence were seen with attendance at two visits (HR 0.27, 95% CI 0.13 to 0.56) and three or more visits (HR 0.18, 95% CI 0.05 to 0.58) (Table 4).

Being aged ≥ 55 years and having an incomplete colonoscopy or colonoscopy of unknown completeness, an adenoma with tubulovillous or villous histology or proximal polyps at baseline were independent risk factors for CRC (see Table 4).

The intermediate-risk group

In the intermediate-risk group, 246 CRCs were diagnosed during 111,270 person-years of follow-up (median 9.1 years, IQR 6.6–12.4 years), giving an incidence rate of 221 per 100,000 person-years (95% CI 195 to 251 per 100,000 person-years). In multivariable regression analysis, number of surveillance visits, age, year of baseline visit, length of baseline visit, completeness of colonoscopy, adenoma dysplasia and proximal polyps were independently associated with CRC incidence. Adenoma histology was not included in the final multivariable model because the association between adenoma histology and CRC incidence was driven by the unknown histology category. Adjusting for the other factors, CRC incidence was 41% lower with attendance at one surveillance visit than with none (HR 0.59, 95% CI 0.43 to 0.81). Incidence rates did not fall much further with attendance at a second surveillance visit (HR 0.56, 95% CI 0.36 to 0.85), but fell again with attendance at three or more visits (HR 0.44, 95% CI 0.26 to 0.77) (Table 5).

Independent risk factors for CRC included age ≥ 65 years, having a baseline visit before 2000 or a baseline visit that spanned between 2 days and 3 months or ≥ 6 months, and having an incomplete colonoscopy or colonoscopy of unknown completeness, an adenoma with high-grade dysplasia or proximal polyps at baseline (see Table 5).

The high-risk group

In the high-risk group, 84 CRCs were diagnosed during 22,961 person-years of follow-up (median 8.4 years, IQR 5.7–11.2 years), giving an incidence rate of 366 per 100,000 person-years (95% CI 295 to 453 per 100,000 person-years). In multivariable regression analysis, number of surveillance visits, completeness of colonoscopy and adenoma dysplasia were independently associated with CRC incidence. Adjusting for these factors, one surveillance visit was associated with a halving of CRC incidence (HR 0.49, 95% CI 0.29 to 0.82), compared with no surveillance. Attendance at subsequent surveillance visits was associated with further reductions in CRC incidence (HR 0.30, 95% CI 0.15 to 0.62 for two visits and HR 0.29, 95% CI 0.11 to 0.73 for three or more visits) (Table 6).

Independent risk factors for CRC included age (≥ 75 years) and having an incomplete colonoscopy, colonoscopy of unknown completeness or an adenoma with high-grade dysplasia at baseline (see Table 6).

In each of the three risk groups there were no significant interactions between the number of surveillance visits and age or sex. The results from the interaction analyses are presented in Appendix 2, Table 26.

The low-risk group

The higher-risk subgroup of the low-risk group comprised patients who had an incomplete colonoscopy or colonoscopy of unknown completeness, a tubulovillous or villous adenoma, or proximal polyps at baseline (n = 9166, 64%). Patients without any of these factors were assigned to the lower-risk subgroup (n = 5235, 36%). Patients in the higher-risk subgroup were older and more likely to have had their baseline visit before 2005 than those in the lower-risk subgroup. Higher-risk patients had more surveillance than lower-risk patients (see Appendix 2, Table 27).

The intermediate-risk group

The higher-risk subgroup of the intermediate-risk group comprised patients who had an incomplete colonoscopy or colonoscopy of unknown completeness, an adenoma with high-grade dysplasia or proximal polyps at baseline (n = 7114, 60%). The lower-risk subgroup comprised patients who had none of these factors (n = 4738, 40%). Patients in the higher-risk subgroup were older, more likely to have had their baseline visit before 2005 and attended more surveillance visits than patients in the lower-risk subgroup (see Appendix 2, Table 27).

The high-risk group

The higher-risk subgroup of the high-risk group comprised patients who had an incomplete colonoscopy or colonoscopy of unknown completeness, or an adenoma with high-grade dysplasia at baseline (n = 902, 33%). Patients without any of these factors were assigned to the lower-risk subgroup (n = 1817, 67%). Patients in the two subgroups were comparable in terms of sex, age and number of surveillance visits. Higher-risk patients were more likely to have had their baseline visit before 2005 than lower-risk patients (see Appendix 2, Table 27).

The low-risk group

Among low-risk patients, attendance at one surveillance visit was associated with reduced CRC incidence in the higher-risk subgroup (HR 0.52, 95% CI 0.35 to 0.78) and even greater reductions were seen with attendance at two or more visits (HR 0.18, 95% CI 0.08 to 0.37). By comparison, in the lower-risk subgroup, surveillance was not associated with significantly reduced CRC incidence rates (HR 0.54, 95% CI 0.25 to 1.20, for one visit and HR 0.42, 95% CI 0.12 to 1.48, for two or more visits). However, it is important to note that the HR estimates for the lower-risk subgroup have wide 95% CIs, as few CRC cases occurred in each stratum (Table 7).

The intermediate-risk group

Among intermediate-risk patients, attendance at one surveillance visit was associated with reduced CRC incidence in the higher-risk subgroup (HR 0.53, 95% CI 0.37 to 0.76) and CRC incidence was even lower with attendance at two or more visits (HR 0.42, 95% CI 0.27 to 0.65). In the lower-risk subgroup of intermediate-risk patients, surveillance was not associated with reduced CRC incidence rates (HR 0.66, 95% CI 0.35 to 1.23 for one visit and HR 0.63, 95% CI 0.31 to 1.29 for two or more visits). As above, however, the HR estimates for the lower-risk subgroup are imprecise, as few CRC cases occurred in each stratum (see Table 7).

The high-risk group

Among high-risk patients, attendance at one surveillance visit was associated with reduced CRC incidence in the higher-risk subgroup (HR 0.41, 95% CI 0.20 to 0.82). Attendance at subsequent visits was associated with further reductions in CRC incidence in this higher-risk subgroup (HR 0.21, 95% CI 0.08 to 0.51 for two or more visits). By contrast, surveillance was not associated with reduced CRC incidence rates in the lower-risk subgroup of high-risk patients (HR 0.66, 95% CI 0.31 to 1.41 for one visit and HR 0.49, 95% CI 0.20 to 1.18 for two visits). However, estimates for the lower-risk subgroup again lack precision (see Table 7).

The low-risk group

Without surveillance, cumulative CRC incidence at 10 years was 1.7% (95% CI 1.4% to 2.1%) in the low-risk group overall, but differed significantly between the lower-risk subgroup, at 1.2% (95% CI 0.8% to 1.7%), and the higher-risk subgroup, at 2.1% (95% CI 1.7% to 2.6%). After one surveillance visit, cumulative CRC incidence at 10 years was 1.5% (95% CI 1.0% to 2.3%) in the low-risk group overall and was no longer significantly different in the lower-risk subgroup (0.6%, 95% CI 0.2% to 1.4%) when compared with higher-risk subgroup (2.0%, 95% CI 1.3% to 3.0%) (Table 8 and Figure 2).

FIGURE 2.

Kaplan–Meier estimates of cumulative CRC incidence after baseline in the low-risk group. (a) Cumulative CRC incidence after baseline with no surveillance (censoring at first surveillance) in the whole low-risk group; (b) cumulative CRC incidence after baseline with no surveillance (censoring at first surveillance) in the higher- and lower-risk subgroups; (c) cumulative CRC incidence with a single surveillance visit (censoring at second surveillance) in the whole low-risk group; (d) cumulative CRC incidence with a single surveillance visit (censoring at second surveillance) in the higher- and lower-risk subgroups; (e) cumulative CRC incidence with two or more surveillance visits (censoring at end of follow-up) in the whole low-risk group; and (f) cumulative CRC incidence with two or more surveillance visits (censoring at end of follow-up) in the higher- and lower-risk subgroups.

The intermediate-risk group

Without surveillance, cumulative CRC incidence at 10 years was 2.6% (95% CI 2.1% to 3.3%) in the intermediate-risk group overall, but differed significantly between the lower-risk subgroup, at 1.3% (95% CI 0.8% to 2.1%), and the higher-risk subgroup, at 3.7% (95% CI 2.9% to 4.7%). After one surveillance visit, cumulative CRC incidence at 10 years was 2.6% (95% CI 1.9% to 3.6%) in the intermediate-risk group overall, still differing significantly between the lower-risk subgroup (1.9%, 95% CI 1.0% to 3.4%) and the higher-risk subgroup (3.1%, 95% CI 2.2% to 4.5%) (Figure 3; see also Table 8).

FIGURE 3.

Kaplan–Meier estimates of cumulative CRC incidence after baseline in the intermediate-risk group. (a) Cumulative CRC incidence after baseline with no surveillance (censoring at first surveillance) in the whole intermediate-risk group; (b) cumulative CRC incidence after baseline with no surveillance (censoring at first surveillance) in the higher- and lower-risk subgroups; (c) cumulative CRC incidence with a single surveillance visit (censoring at second surveillance) in the whole intermediate-risk group; (d) cumulative CRC incidence with a single surveillance visit (censoring at second surveillance) in the higher- and lower-risk subgroups; (e) cumulative CRC incidence with two or more surveillance visits (censoring at end of follow-up) in the whole intermediate-risk group; and (f) cumulative CRC incidence with two or more surveillance visits (censoring at end of follow-up) in the higher- and lower-risk subgroups.

The high-risk group

Without surveillance, cumulative CRC incidence at 10 years was 5.7% (95% CI 4.0% to 8.3%) in the high-risk group overall, differing significantly between the lower-risk subgroup, at 3.8% (95% CI 2.1% to 6.8%), and the higher-risk subgroup, at 9.9% (95% CI 6.2% to 15.7%). After one surveillance visit, CRC incidence at 10 years was 5.6% (95% CI 3.1% to 9.8%) in the high-risk group overall and was no longer significantly different in the lower-risk subgroup (4.4%, 95% CI 1.8% to 10.6%) when compared with the higher-risk subgroup (7.8%, 95% CI 3.8% to 15.4%) (Figure 4; see also Table 8).

FIGURE 4.

Kaplan–Meier estimates of cumulative CRC incidence after baseline in the high-risk group. (a) Cumulative CRC incidence after baseline with no surveillance (censoring at first surveillance) in the whole high-risk group; (b) cumulative CRC incidence after baseline with no surveillance (censoring at first surveillance) in the higher- and lower-risk subgroups; (c) cumulative CRC incidence with a single surveillance visit (censoring at second surveillance) in the whole high-risk group; (d) cumulative CRC incidence with a single surveillance visit (censoring at second surveillance) in the higher- and lower-risk subgroups; (e) cumulative CRC incidence with two or more surveillance visits (censoring at end of follow-up) in the whole high-risk group; and (f) cumulative CRC incidence with two or more surveillance visits (censoring at end of follow-up) in the higher- and lower-risk subgroups.

The low-risk group

Compared with the general population, CRC incidence in the absence of surveillance was not significantly different in the low-risk group overall (SIR 0.86, 95% CI 0.73 to 1.02) or in the higher-risk subgroup (SIR 1.07, 95% CI 0.88 to 1.28), but was lower in the lower-risk subgroup (SIR 0.51, 95% CI 0.35 to 0.73). After one surveillance visit, CRC incidence was significantly lower in both subgroups than in the general population and SIRs were 0.38 (95% CI 0.16 to 0.74) and 0.70 (95% CI 0.48 to 0.98) for the lower- and higher-risk subgroups, respectively (see Table 8).

The intermediate-risk group

Without surveillance, CRC incidence was not significantly different in the intermediate-risk group overall (SIR 1.16, 95% CI 0.97 to 1.37) compared with the general population, but was lower in the lower-risk subgroup (SIR 0.70, 95% CI 0.48 to 0.99) and higher in the higher-risk subgroup (SIR 1.46, 95% CI 1.19 to 1.78). After one surveillance visit, CRC incidence was no longer higher in the higher-risk subgroup than in the general population (SIR 1.00, 95% CI 0.73 to 1.33) (see Table 8).

The high-risk group

Without surveillance, CRC incidence was higher in the high-risk group overall (SIR 1.91, 95% CI 1.39 to 2.56) and in the higher-risk subgroup (SIR 3.55, 95% CI 2.34 to 5.17) than in the general population, but was not significantly different in the lower-risk subgroup (SIR 1.10, 95% CI 0.64 to 1.76). After first surveillance, CRC incidence was not significantly different in the high-risk group overall (SIR 1.34, 95% CI 0.86 to 1.99) than in the general population, but remained higher in the higher-risk subgroup (SIR 1.97, 95% CI 1.02 to 3.44). Following a second surveillance visit, CRC incidence was no longer higher in the higher-risk subgroup than in the general population (SIR 1.02, 95% CI 0.41 to 2.09) (see Table 8). It is worth noting that the estimates from our analyses of the high-risk group have low precision because of the small number of CRC cases.

The low-risk group

In the whole low-risk group, CRC incidence was lower with attendance at two or more surveillance visits than with one visit (HR 0.31, 95% CI 0.15 to 0.63). This pattern was also observed in the higher-risk subgroup (HR 0.25, 95% CI 0.11 to 0.57). In the lower-risk subgroup, attendance at two or more surveillance visits was not associated with a reduction in CRC incidence compared with attendance at one visit (HR 0.74, 95% CI 0.14 to 3.81). However, the HR estimate for the lower-risk subgroup has a wide 95% CI, as there were only 11 CRC cases (Table 9).

The intermediate-risk group

In the whole intermediate-risk group, attending two or more surveillance visits was associated with a reduction in CRC incidence compared with a single visit (HR 0.60, 95% CI 0.40 to 0.91). This was also observed in the higher-risk subgroup (HR 0.56, 95% CI 0.34 to 0.91), but not in the lower-risk subgroup (HR 0.66, 95% CI 0.30 to 1.46). However, as before, the HR estimate for the lower-risk subgroup is imprecise as there were few CRC cases (see Table 9).

The high-risk group

In the whole high-risk group, CRC incidence was lower with attendance at two or more surveillance visits than with one visit (HR 0.43, 95% CI 0.21 to 0.86). This was also observed in the higher-risk subgroup (HR 0.36, 95% CI 0.13 to 0.99). In the lower-risk subgroup, CRC incidence was not significantly lower with two or more surveillance visits than with one (HR 0.50, 95% CI 0.19 to 1.30), although the HR estimate lacks precision (see Table 9).

Detection rates of advanced adenomas and colorectal cancer at first surveillance, considering all possible surveillance intervals

Detection rates of advanced adenomas and colorectal cancer at first surveillance, applying our surveillance interval length cut-off points

For our main analyses of findings at first surveillance we excluded patients whose first surveillance examination was performed after our selected interval length cut-off points (i.e. 8.5 years for the low-risk group, 6.5 years for the intermediate-risk group and 3.75 years for the high-risk group). We examined the effect of interval length on AA and CRC detection rates at first surveillance in the risk groups overall and by risk subgroup.

Patients without a complete baseline colonoscopy

Our first sensitivity analysis involved the exclusion of patients who did not have a complete baseline colonoscopy (low-risk group, n = 2682; intermediate-risk group, n = 2885; high-risk group, n = 365). This had little effect on the results for the long-term CRC incidence analyses (see Appendix 2, Tables 28–32), although adenoma histology, adenoma dysplasia and age were no longer independently associated with CRC incidence in the low-, intermediate- and high-risk groups, respectively (see Appendix 2, Tables 28–30).

We also performed this sensitivity analysis for the detection rates of AAs and CRC at first surveillance by interval length and risk subgroup (see Appendix 2, Table 33) and the results were similar to those of the main analysis. A minor difference was that in the high-risk group there was no longer a significant trend of increasing CRC detection with increasing interval in the higher-risk subgroup, nor a significant difference in the odds of CRC detection between the two risk subgroups. However, the number of CRC cases in the high-risk group was low in this analysis.

Incompletely resected baseline lesions

We also performed a sensitivity analysis of the criteria used to decide whether or not a CRC had arisen from an incompletely resected baseline lesion, which we excluded from analyses. In the main analysis we excluded cancers that were found in the same or adjacent colonic segment to a baseline adenoma that was ≥ 15 mm in size and seen on two or more occasions within 5 years before cancer diagnosis (low-risk group, n = 0; intermediate-risk group, n = 38; high-risk group, n = 12). In the sensitivity analysis we excluded some additional cancers that met some but not all the criteria above, but that we thought were likely to have arisen from an incompletely resected baseline lesion (low-risk group, n = 6; intermediate-risk group, n = 29; high-risk group, n = 7).

These additional exclusions did not materially alter the results (see Appendix 2, Tables 34–37). In the analyses of long-term CRC incidence by baseline characteristics there were slight changes to the associated p-values for some variables, such that they crossed to the opposite side of the 0.05 significance threshold, although the HR point estimates did not change substantially (see Appendix 2, Tables 34–36). In the intermediate-risk group there were changes in significance for sex, year of baseline visit and length of baseline visit, although no such changes were seen for any of the variables used to classify the risk subgroups.

In addition, in all three risk groups the estimates for the comparison of CRC incidence in the presence and absence of surveillance were similar in the sensitivity analysis and the main analysis (see Appendix 2, Tables 34–36). We also observed little change in results when we performed this sensitivity analysis for the detection rates of AAs and CRC at first surveillance by interval and risk subgroup (see Appendix 2, Table 37).

Higher-risk classification criteria for the intermediate-risk group

In our sensitivity analysis of the risk classification criteria for intermediate-risk patients we additionally included poor bowel preparation and adenomas ≥ 20 mm in size in the classification of higher risk. As a result, the proportion of patients classified as higher risk increased from 60% to 74% in our analyses of long-term CRC incidence. This did not have much impact on CRC incidence rates, estimates for the comparison of CRC incidence in the presence and absence of surveillance or SIRs (see Appendix 2, Tables 38 and 39). In our analyses of findings at first surveillance the proportion of patients classified as higher risk increased from 61% to 76% and results by interval length within each risk subgroup were not materially altered (see Appendix 2, Table 40).

Surveillance interval length cut-off point for the low-risk group

In a fourth sensitivity analysis we changed our interval length cut-off point for the low-risk group from 8.5 years to 6.5 years. With this change in cut-off point there was still a significant trend of increasing detection of AAs at first surveillance with increasing interval length in both risk subgroups. Detection rates of AAs and CRC were also still significantly higher in the higher-risk subgroup than in the lower-risk subgroup. However, there was no longer a significant trend in CRC detection rates with interval length in either risk subgroup (see Appendix 2, Table 41).

Hospital attended

For our final sensitivity analysis we also considered the hospital that patients attended as a potential confounding variable. In the analyses of long-term CRC incidence the hospital attended was independently associated with incidence rates in the low- and intermediate-risk groups, but not in the high-risk group. When we additionally adjusted for hospital the multivariable HRs for the other variables changed very little, although adenoma histology was no longer significant in the low-risk group (see Appendix 2, Table 42).

Form of evaluation

The economic analysis consists of both a within-study analysis using individual patient-level data recorded on the study database and a lifetime analysis using a Markov model. The study database provided resource use information on the number and type of colonic examinations. Both analyses included all 28,972 patients included in the clinical analysis (i.e. 14,401 low-risk patients, 11,852 intermediate-risk patients and 2719 high-risk patients).

For the within-study analysis an annual rate of total cost per person-year was calculated for each risk subgroup across a median follow-up of 9.3 years. As the study database did not include any information on QoL, it was not possible to estimate QALYs directly from the study data. The within-study analysis used the diagnosis of CRC as the main outcome measure, with cost-effectiveness expressed in terms of the incremental cost per CRC prevented.

An extrapolation model was then used to estimate the cost-effectiveness of each surveillance strategy over the lifetime of the cohort. Surveillance costs and transition probabilities for each risk subgroup were estimated using patient-level data on the study database. For this analysis to measure outcomes using QALYs, QoL data were obtained from the PHE PROMs survey of CRC patients known to the National Cancer Registration and Analysis Service37 and combined with survival data from the Northern and Yorkshire Cancer Registry and Information Service. 42

Estimation of costs

Table 17 shows the unit costs used in our analyses and the sources of these costs. Costs were applied to the resource use associated with both adenoma surveillance and cancer treatment. The analysis was undertaken from the perspective of the NHS, with costs reported in GBP using 2017/18 prices. Unit costs for surveillance procedures were taken from the NHS national schedule of reference costs for 2017–18. 33 Estimates for the lifetime cost of CRC treatment were taken from published estimates from a whole disease model of CRC. 44 These costs were inflated from 2012/13 prices to 2017/18 prices using the gross domestic product deflator.

The definition of surveillance visits used in previous chapters is maintained in this chapter. Visits were costed by applying unit costs to each colonic examination. Costs were different for colonoscopies, flexible sigmoidoscopies and rigid sigmoidoscopies, and varied depending on whether the examination was diagnostic or therapeutic. The distinction between a diagnostic and a therapeutic examination was made using data on whether or not a polyp was removed or whether there was an associated biopsy or pathology report.

Within-study analysis

The within-study analysis compared patients who underwent surveillance with those who did not for each of the six risk subgroups (i.e. for the lower- and higher-risk subgroups within each of the low-, intermediate- and high-risk groups). Cost-effectiveness was measured in terms of the incremental cost per CRC prevented by adopting surveillance compared with no surveillance. This follows the approach of a recent cost-effectiveness study of post-polypectomy surveillance. 45

Total annual costs and CRC incidence rates were calculated for each of the three main risk groups and for the lower- and higher-risk subgroups within each risk group. Poisson models were used to estimate annual CRC incidence rates for each risk group across varying exposure time. Differences in CRC incidence rates were compared using the Wald test, with SEs being combined using the delta method. We calculated ICERs as the ratio between the difference in costs and the difference in CRC incidence rates.

Missing data were infrequent in the data set, with the exception of CRC staging information, which was missing for 32% of CRC cases. Logit models indicated that these missing data were significantly positively associated with age at baseline and negatively associated with the date of the surveillance visit (data not shown). Missing CRC staging data were handled by means of multiple imputation using an ordered logit model. The procedure was repeated to produce 40 imputed data sets, with Rubin’s rule used to summarise across imputations. 46 In the main analysis, total costs were then estimated based on the imputed data. Ten patients in the data were untraceable. These patients were treated as having had no surveillance visits and having not developed CRC. Both costs and number of CRC cases were discounted at a rate of 3.5% per year. The analysis was conducted in Stata^®/SE 14 (StataCorp LP, College Station, TX, USA).

Extrapolation model

We designed a multistate Markov model to extrapolate the results from the within-study analysis over a lifetime horizon. Figure 5 shows the structure of the model. The model consisted of eight states: (1) no surveillance visits, (2) a positive count of surveillance visits, (3) Dukes’ stage A CRC, (4) Dukes’ stage B CRC, (5) Dukes’ stage C CRC, (6) Dukes’ stage D CRC, (7) death from CRC and (8) death from other causes. Time-homogeneous transition probabilities were estimated for each of the three main risk groups using state and time data from the main study database and using the msm package in R (The R Foundation for Statistical Computing, Vienna, Austria). SEs for each set of transition probabilities were calculated using an assumed multivariate normal distribution of the maximum likelihood estimates and covariance matrix.

FIGURE 5.

Markov model for our lifetime economic analysis.

All patients begin the model in the ‘no visits’ state, having attended a baseline colonoscopy but having had no surveillance visits. In each cycle patients have a probability of attending a surveillance visit that is dependent on which risk group they are in. The patients also face a probability of developing CRC and having it diagnosed at Dukes’ stage A, B, C or D. A transition to the ‘visits’ state reduces the patient’s risk of CRC and this benefit is assumed to last for the remaining length of the model. The transition probabilities between the ‘no visits’ state and the ‘visits’ state, as well as the transitions to each of the cancer states, were estimated separately for each risk group.

The numbers of surveillance visits and intervals between them as observed within each risk subgroup in the study data were used to estimate a constant annual probability for each subgroup. The model does not explicitly model the impact of the number of surveillance visits as the surveillance strategies being compared are defined in terms of whether or not surveillance occurred. However, the impact of the number of surveillance visits on CRC cases diagnosed is captured in the probability of transitioning from the ‘visits’ state to each of the four cancer states.

Missing CRC staging data were handled using the imputed data from the within-study analysis. Transition probabilities were calculated for each of the 40 imputed data sets, with SEs estimated by combining the estimates using Rubin’s rule. 46 The sensitivity of the results to the use of the imputed CRC data were assessed by collapsing the model to a single cancer state for each of the risk subgroups.

A full set of transition probabilities and SEs could not be estimated for the higher-risk subgroup of the high-risk group because of a failure of model convergence. Therefore, a simplified version of the model was used for this group, in which the four cancer states were collapsed to a single state. QoL estimates, CRC treatment costs and CRC survival rates, which are separately estimated for each cancer state, were collapsed to mean values for this subgroup.

The probability of death from any cause, QoL estimates and costs associated with CRC are dependent on the age of the cohort. These parameters required an initial age to be specified for the model. The initial age used for the base-case analysis was 60 years, as this was the mean age of patients in the hospital cohort.

The main clinical study was not powered to identify differences in CRC-related mortality by stage and there were relatively few CRC deaths in each category when stratifying by stage, which would make extrapolation based on these data unreliable. Therefore, the probabilities of survival of CRC were sourced from the 2013 National Bowel Cancer Audit Annual Report. 35 The probability of death from any cause was taken from national lifetables published by the Office for National Statistics. 36

Resource use and costs

The extrapolation model includes costs for colonoscopy and lifetime costs associated with the diagnosis and treatment of CRC. A separate cost for adenoma surveillance was entered into the model for each risk subgroup. This cost was derived by estimating the mean annual cost of surveillance per patient during the within-study period and then applying that cost over the modelled duration of surveillance, from a starting age of 60 years through to 75 years, the upper age limit in the 2002 UK-ASG. 7 The lifetime cost of CRC was applied only to the first cycle in which CRC was diagnosed.

Quality of life

Quality-of-life data were not collected in the main clinical study. Therefore, the primary outcome measure in the within-study analysis was the incidence of CRC. The extrapolation model used estimated mean EQ-5D scores from a one-off study of CRC patients by a PHE PROMs survey. 37 This study collected QoL data from 21,802 patients who were diagnosed with CRC in England from 2010 to 2011 and were alive between 12 and 36 months after diagnosis, when they were sent a questionnaire including the EuroQol-5 Dimensions, five-level version. Responses were then scored using the ‘crosswalk’ mapping function developed by van Hout et al. 47 This allowed for comparison with the utility estimates derived from the EuroQol-5 Dimensions, three-level version (EQ-5D-3L), which was used for the non-cancer health states. The EQ-5D-3L scores for the non-cancer health states came from Ara and Brazier,38 who reported EQ-5D-3L scores from pooled responses in the Health Survey for England from people without cancer.

The Ara and Brazier38 scores and the estimates from the PROMs survey37 were combined with estimated survival from the lifetime model to calculate QALYs. The use of the PROMs survey for QoL data differs from recent studies examining the cost-effectiveness of CRC screening or surveillance strategies,23,43 which have mainly used estimates reported in Ara and Brazier. 38 The estimates from the PROMs survey37 were preferred over those from Ara and Brazier38 as the former is the largest UK survey of QoL for CRC patients currently available, and provides estimates by both CRC stage and age. Whyte et al. 48 estimated EQ-5D-3L scores by age and estimated CRC stage; however, these estimates are based on all cancer patients included in the Health Survey for England rather than only CRC patients. Therefore, the PROMs data were deemed more representative of the patients in our study cohort. The effect of using the estimates from Whyte et al. 48 on our model results was assessed in a sensitivity analysis. 48

Future QALYs and costs were discounted to present values at an annual rate of 3.5%. ICERs were then calculated as the ratio between the mean difference in QALYs and the mean difference in costs. Cost-effectiveness was evaluated assuming a willingness-to-pay threshold of £20,000 per QALY. 49

Sensitivity analyses

Uncertainty in the model estimates was characterised using a deterministic sensitivity analysis (DSA) and probabilistic sensitivity analysis (PSA). The DSA assessed the sensitivity of the model estimates to variation in individual parameters, whereas the PSA aimed to estimate the joint effect of uncertainty in all the parameters. For the DSA each parameter was both increased and decreased by 25% of its baseline value. The impact of these changes was measured in terms of the effect on the estimated ICER for each surveillance strategy. A sensitivity analysis was also used to assess the impact of using different strategies to handle the missing CRC staging data.

In the PSA all transition probabilities, costs and QoL estimates were varied, except for the survival probabilities. The PSA was carried out by sampling 2000 sets of the model parameters drawn at random from appropriate statistical distributions. Uncertainty around each cost estimate was characterised using a gamma distribution, whereas beta distributions were fitted to both the transition probabilities and QoL estimates. These replications were used to plot the cost-effectiveness plane50 and to construct cost-effectiveness acceptability curves that show the likelihood that the intervention is cost-effective as the willingness-to-pay changes. 51

Table 18 shows the estimated transition probabilities, costs and QoL estimates used in the model. The transition probabilities between the ‘no visit’ and ‘visit’ states differ between risk groups, with the lowest probability estimated for the lower-risk subgroup of the low-risk group and the highest probability estimated for the higher-risk subgroup of the high-risk group. For the whole low-risk group and the higher-risk subgroup of intermediate-risk patients the transition probabilities to the different CRC states were mostly lower with surveillance than without. In contrast, for the lower-risk subgroup of intermediate-risk patients and the lower-risk subgroup of high-risk patients the transition probabilities to CRC were mostly higher with surveillance than without. Mean EQ-5D scores were higher for patients with CRC than for people in the same age group without cancer.

The below list summarises the main modelling assumptions made in the lifetime analysis:

• The level of surveillance recorded in the study data set reflects practice under the 2002 UK-ASG. 7

• The level of surveillance offered to each risk group is independent of the level of surveillance offered to the other risk groups.

• Transition probabilities are time homogeneous.

• Patients who did not attend surveillance are representative of patients who attended surveillance if their surveillance was withdrawn.

• There is no age cut-off point for surveillance in either of the considered surveillance strategies.

• There were no changes to the BCSP during the modelled period.

• The probability of a bleed occurring during a surveillance procedure is independent of polypectomy being performed.

• The probability of a bleed or bowel perforation occurring during a surveillance procedure is independent of risk group.

• The costs and QoL estimates associated with surveillance can be captured by explicitly modelling the growth of polyps.

Although we assumed that the probability of a bleed occurring during a surveillance procedure is independent of polypectomy being performed, polypectomy does in fact increase the risk of a bleed. 52 However, a simplification of the model is that it does not distinguish between a polypectomy being performed or not. Therefore, when applying the probability of a bleed it was necessary that the probability was independent of a polypectomy being performed. To achieve this we used an average across all colonoscopies performed sourced from Rutter et al. 52

Within-study analysis

Of the 525 CRCs in the study, 167 (32%) were missing CRC staging data. The proportion of missing CRC staging data varied between the risk groups. The highest proportion was 35% in the higher-risk subgroup of low-risk patients, whereas the lowest proportion was 24% in the lower-risk subgroup of intermediate-risk patients. However, the risk group was not a significant predictor of missing CRC staging data (data not shown).

Lifetime analysis

A comparison of observed and predicted percentage in each state from a single imputed data set showed that the model fitted the data reasonably well (see Appendix 3, Figures 9–14). The results from our lifetime model are shown in Table 22. For each of the three main risk groups ICERs were lower in the higher- than in the lower-risk subgroup.