The cost-effectiveness of domiciliary non-invasive ventilation in patients with end-stage chronic obstructive pulmonary disease: a systematic review and economic evaluation

Diagnosis

Airflow obstruction is diagnosed by spirometry, performed after use of a bronchodilator. Spirometry generates a forced expiratory volume in 1 second (FEV₁) and a forced vital capacity (FVC); it is the ratio between these two figures that defines airflow obstruction. Although definitions of airflow obstruction are inconsistent and controversial,6 National Institute for Health and Care Excellence (NICE) guidance currently defines airflow obstruction as FEV₁/FVC< 0.7. 4 This is one of two common definitions; the other is to define it as any value of FEV₁/FVC that is less than the lower limit of normal for the patient’s age. 6

Severity

Severity of airflow obstruction is graded using categories of FEV₁ as a percentage of predicted normal values of a healthy reference population. 1,4 Table 1 outlines some severity categories. The definitions of different grading categories vary somewhat and have changed over time. For example, the Global Initiative for Chronic Obstructive Lung Disease (GOLD)1 has proposed a revised COPD grading system that incorporates symptoms and HRQoL into the severity categories.

Most research studies evaluating treatments use FEV₁ % predicted to select and describe patients. FEV₁ is also often used as an outcome measure to describe prognosis of patients, as are clinical measures such as dyspnoea and exacerbations, HRQoL and health-service utilisation (e.g. hospital admissions1).

Severity of airflow obstruction does not necessarily reflect either the level of disability experienced or the frequency of exacerbation, and composite measures to capture the global impact of the disease have been proposed. 1,7 However, they are not yet widely used as the basis for treatment decisions.

In more severe disease, respiratory failure can occur [hypoxia with resting pO₂ (partial pressure of oxygen) < 8 kPa]. Respiratory failure may be either type 1, in which arterial carbon dioxide is normal or low, or type 2, in which the patient is hypercapnic (elevated arterial CO₂). It is not yet clear which factors predispose to the different types of respiratory failure or how quickly it will develop or progress. Type 2 respiratory failure can also result in acidosis of the circulating blood; this is known as acute hypercapnic respiratory failure (HRF). Acidosis is not present in the stable state, as it is compensated for by renal mechanisms of acid–base balance.

End-stage COPD could be defined in several ways; classically, it would concern those patients in the terminal stage of their disease who were likely to die within months, a situation which is not always clear. Attempts are being made to define the disease trajectory at the end of life, but it remains hard to predict. 8 Alternatively, patients with end-stage COPD might be defined as those who have developed chronic respiratory failure and remain symptomatic on maximal therapy, with no hope of cure. This is a significant population, who could potentially be stabilised for years. For the purposes of this report, the latter, broader definition prevails, as this encompasses the wide population for whom domiciliary non-invasive ventilation (NIV) has been considered in some studies and in practice.

Exacerbations

Exacerbations or ‘flare-ups’ of COPD occur in approximately 50–60% of moderate/severe COPD patients per year. 2,9 An exacerbation can be defined as an acute event characterised by a worsening of the patient’s respiratory symptoms that is beyond normal day-to-day variations and leads to a change in medication. 1 The most common causes are respiratory tract infections, either viral or bacterial, although other factors, including pollution,10 can precipitate them. Exacerbations play a very important role in COPD, as they are a key cause of increased morbidity, mortality and poor health status and place a considerable burden on the health-care system. 11

It is estimated that approximately 75% of exacerbations can be managed in the community with antibiotics, corticosteroids and bronchodilators. 12 Half of exacerbations treated in the community result in patients recovering to their baseline level in about 7 days; however, in 14% of such events patients do not reach their baseline level after 35 days, and a few never return to baseline. 13

In general, a severe exacerbation is defined as one that results in hospitalisation (Table 2). Approximately 15% of COPD patients per year have exacerbations severe enough to lead to hospital admission, with a median length of stay in the UK of 8 days. This contributes to over half of the total direct costs of COPD to the NHS. 3,15 Between 10% and 25% of patients admitted with HRF caused by COPD die in hospital. 16 Readmission for an exacerbation within 3 months is high, at over 30%, as is 30-day mortality. In those requiring artificial ventilation in hospital (intubation or NIV), mortality may be as high as 40% at 1 year after discharge, and the all-cause mortality 3 years after hospitalisation is higher still. 16 It is therefore evident that hypercapnic COPD patients, and those who have previously used ventilation in hospital, may have a poor prognosis.

Exacerbations have an adverse effect on a patient’s quality of life (QoL). 11 It may take many weeks for the symptoms to abate and lung function to recover; they have also been shown to lead to a rapid overall decline in lung function. 17 Some patients appear to suffer from frequent exacerbations while others do not. Those reporting two or more exacerbations of COPD per year are classified as ‘frequent exacerbators’. 2

Exacerbations are important outcome measures in COPD and a reduction in exacerbation frequency is an important target to achieve for any intervention.

Quality of life

Health-related QoL is known to be impaired by COPD, even at relatively early stages,18 and gets worse with more severe disease and with exacerbations. 19 Many tools are available to monitor HRQoL in COPD, and have been reviewed elsewhere;20 they may be generic (i.e. used in any health problem), respiratory-specific or COPD-specific (e.g. COPD assessment test). 21 Tools commonly used in clinical trials include the St George’s Respiratory Questionnaire (SGRQ)22 and the Chronic Respiratory Questionnaire (CRQ). As with all questionnaires, it is possible that scores vary between the tools used; indeed, this has been the case in COPD when comparing a generic tool [Short Form questionnaire-36 items (SF-36)], the SGRQ and the CRQ. 23 Other tools include the Severe Respiratory Insufficiency Questionnaire (SRI) and Maugeri Foundation Respiratory Failure questionnaire (MRF-28). 24

Management of chronic obstructive pulmonary disease

Pharmacotherapy in stable COPD is mainly directed at airflow obstruction and inflammation. It is recommended that short- and long-acting bronchodilators be used in a stepwise manner, with the addition of inhaled corticosteroids later in the disease. In addition to pharmacotherapy, pulmonary rehabilitation is recommended for breathless patients. 4 This is a programme of exercise and education which has been shown to be very beneficial in COPD patients. 25 Patients may also be taught various self-management techniques, the efficacy of which has recently been reviewed by some of the authors of this report. 26

In general, treatment is aimed at improving lung function, defined by FEV₁, and/or HRQoL and exacerbation frequency. Many drug treatments have also tried to improve longer-term outcomes, such as lung function decline12 and mortality,27 with limited success. In addition to the general treatments described above, long-term oxygen therapy (LTOT) should be considered28 in patients with pO₂ < 7.3 kPa, or < 8 kPa in the presence of cor pulmonale,29 because it improves mortality. Other forms of oxygen use in COPD are more controversial; oxygen use just on walking (ambulatory oxygen) is used in selected patients not on LTOT if their oxygen levels drop on objective testing. 30 Short-burst oxygen therapy, defined as oxygen used outside the context of LTOT or ambulatory oxygen, is not recommended because of lack of benefit. 31 Another treatment that has been proposed for selected COPD patients is domiciliary use of NIV, which is the subject of this review.

Types of non-invasive ventilation

Non-invasive ventilation delivers two different pressures; that is, a different pressure on inspiration [inspiratory positive airway pressure (IPAP)] from expiration [expiratory positive airway pressure (EPAP)]. It differs from continuous positive airway pressure (CPAP), another non-invasive form of respiratory support, which aims to maintain a continuous level of positive airway pressure. The type of NIV is generally described by the way in which the machine is set up, such that it will be either pressure controlled or volume controlled. In pressure-controlled settings, the main descriptors for treatment will be the IPAP and EPAP, although a back-up rate will usually also be set. The mask type may also be described, with most patients using a full-face mask, although nasal-only devices are also available. Once patients start NIV, the parameters can be varied according to their ability to tolerate treatment and their response; for instance, an in-patient NIV protocol might start at relatively low IPAP and titrate upwards until the patient’s blood gases improve, stopping any further elevation in pressure if the patient finds it uncomfortable. While there is no widely accepted definition of low-pressure and high-pressure NIV, there has been work to suggest that degree of IPAP may relate to outcome, and there has been an increasing trend towards the use of higher pressures in more recent studies32,33 compared with older studies. 34,35 The type of NIV most useful to COPD is not clear; neither are the optimal pressure setting and optimisation protocol. It is likely that there will be significant patient variability and individualised protocols will be required.

Domiciliary NIV is usually used overnight during sleep, although hours of use may gradually climb with time until daytime use is also needed.

Use of non-invasive ventilation for chronic obstructive pulmonary disease in hospital settings

There is good evidence that HRF during an acute exacerbation should be treated with NIV;4 a Cochrane review (Ram et al. 200436) of NIV found statistically significant benefits in favour of NIV for a wide range of outcomes, including treatment failure (based on eight trials), risk of intubation (based on 14 trials), length of hospital stay (based on eight trials), complications (based on three studies) and mortality (based on 10 trials).

Use of non-invasive ventilation in stable chronic obstructive pulmonary disease patients

The evidence from NIV use in a hospital setting led to trials of domiciliary NIV, aimed at reducing mortality and readmission rates; this is the subject of this report.

Increasingly employed for longer-term treatment of patients suffering from chronic HRF due to thoracic cage disorders, neuromuscular disorders and various other causes of nocturnal hypoventilation syndrome,37 long-term NIV offers theoretical benefits in certain groups of chronic end-stage COPD patients. Clinically, despite the success of NIV in acute HRF with COPD, survivors continue to suffer from further episodes/exacerbations after discharge. 38 Thus, a utility of domiciliary NIV is thought to lie in preventing recurrent admission to hospital and slowing declining health.

The National COPD strategy consultation in 2010 highlighted NIV as an area of COPD therapy that warranted review and/or further research. 39

Population

The National Institute for Health Research (NIHR) call relating to this project (11/27/01) specified ‘adult patients with stable end-stage COPD plus chronic HRF, who have required assisted ventilation (whether invasive or non-invasive) during an exacerbation or who are hypercapnic or acidotic on LTOT.’

The following points are worth considering:

• This specification includes a number of types of patient, that is:

  • stable patients who have required assisted ventilation during an exacerbation

  • stable patients who are hypercapnic on LTOT (regardless of whether or not they have required assisted ventilation during an exacerbation)

  • patients who are acidotic on LTOT.

• ‘Acidotic on LTOT’ implies the absence of stable disease, as it usually requires acute treatment in hospital.

• The determination of whether or not a patient has stable disease and/or end-stage disease is, to some extent, subjective.

In order to capture all relevant groups, studies with both stable patients and post-hospital (exacerbation) patients were eligible for inclusion into separate analyses.

This report therefore necessarily considers a broad definition of patients, that is any adult patient with COPD with or without HRF (however defined), and includes any patients recently discharged from hospital (following an exacerbation) or more stable patients, without restriction to a specific disease severity.

Setting

The NIHR brief specified a community setting and this was taken to mean any setting where NIV was not used in an acute setting or a research setting. Therefore, evidence was not considered when NIV was given in hospital or immediately before hospital during an exacerbation or, for example, when NIV was given during exercise assessment as part of a pulmonary rehabilitation programme. In practice, only studies where patients used NIV in their own home were identified, although other domiciliary settings, such as a care home, would have been eligible.

Intervention

Any form of NIV, whether continuous or intermediate, added to (any form of) usual care was considered. There were no restrictions according to length of daily use.

While the term ‘NIV’ in its broadest sense refers to any type of NIV, as opposed to invasive techniques, in this context it is used to mean a system that delivers two different pressures. The following terms can all mean NIV (i.e. systems with two positive pressures), provided they are non-invasive: bilevel positive airway pressure (BIPAP), non-invasive positive-pressure ventilation or nasal intermittent positive pressure ventilation (both referred to as NIPPV), non-invasive mechanical ventilation, positive pressure ventilation and (nasal) proportional assist ventilation.

Negative pressure ventilation (NPV) is rarely used routinely but also delivers two pressures (although one relates to ambient pressure); studies using NPV would have been eligible for inclusion and would have been considered separately (none was identified). CPAP was excluded, as it delivers constant pressure and is thus a distinct therapy; it is primarily used to treat sleep apnoea.

Comparator

While the NIHR brief did not specify looking at different types of NIV, it is becoming apparent that higher pressures are increasingly being used (see Chapter 1, Types of non-invasive ventilation). Therefore, studies with head-to-head comparisons of different NIV settings (particularly pressure, or pressure vs. volume controlled) were of interest and considered. Other differences include mask type and number of hours of use per day. Length of daily use and a need for longer periods of use, for example during the day as well as during the night may also be related to severity of disease.

Discontinuing domiciliary NIV was not covered by the remit for this report. Therefore it is not considered. However, the availability of evidence on the effect of discontinuing was noted (see Chapter 4, Discontinuation studies).

Outcomes

As the NIHR brief was for an economic evaluation, the primary effectiveness outcomes were considered to be those relating to survival, health-care service utilisation (as a results of exacerbations), patient QoL and adherence to NIV/discontinuations. Secondary outcomes {relating to lung function, blood gases, exercise capacity [e.g. 6-minute walking distance (6MWD)], quality of sleep, activities of daily living and acceptability} were not relevant for the economic model, as cost and utility changes associated with changes in these outcomes cannot be measured. However, in order to present a full clinical picture of the effect of NIV, and to be consistent with previous systematic reviews, data on key secondary outcomes were extracted and presented. A decision was made to focus on lung function (FEV₁, FVC), blood gases [partial pressure of carbon dioxide in the arterial blood (PaCO₂), partial pressure of oxygen in the arterial blood (PaO₂)] and 6MWD. Hydrogen carbonate (HCO₃^–) and pH are to some extent reflected in other blood gas measures, and measures of activity are reflected to some extent in QoL measures and 6MWD. Further, there is a lack of validated questionnaires for activities of daily living and also for sleep quality; hence, data were not extracted for these outcomes.

Study design

Based on an overview of existing clinical effectiveness reviews of RCTs, it was apparent that a lack of large, long-term studies, variations between study methods and physiological or clinical outcomes measured, and a lack of adjustment for clinical variables (such as oxygen use or prior acute NIV use) have limited the conclusions that can be drawn. In particular, RCTs appear to have insufficiently long follow-up periods to capture outcomes relating to survival, long-term HRQoL, exacerbations over the long term, adverse events or adherence rates. Further, the clinical perception is that inclusion criteria for the RCTs may be narrow and restricted to patients with very specific characteristics, calling into question their applicability to a wider population. There may also be clear subgroups within current trials which warrant formal meta-analysis not conducted in prior reviews.

As a further review based on RCTs alone might be insufficient to derive all the necessary (long-term) parameters needed to adequately populate an economic model, non-randomised controlled studies and uncontrolled studies are also considered in this report.

Screening and study selection strategy

Titles (and abstracts, where available) of articles identified by the searches were screened by two reviewers for relevance to the review question using prespecified screening criteria. Hard copies of potentially relevant articles were acquired and assessed against the full inclusion criteria (Table 3) by two reviewers independently. Discrepancy between reviewers was resolved by discussion or by referring to a third reviewer. Where necessary, translation (full/part) of non-English-language articles was undertaken to facilitate this process and subsequent reviewing. The study selection process was illustrated using a PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram. Reference management software (Reference Manager version 11, Thomson ResearchSoft, San Francisco, CA, USA) was used to record reviewer decisions, including reasons for exclusion.

Selection criteria

Primary outcomes

Narrative synthesis of evidence was undertaken for all included (controlled) studies. For primary outcomes, meta-analysis was undertaken in Stata (Version 10, StataCorp LP, College Station, TX, USA) where there was clinical and methodological homogeneity between studies reporting the same outcome and using the same outcome statistic (reported or calculable, see Appendix 3 for more detailed methods). As outlined in the decision problem (see Chapter 1, Decision problem), a distinction was made between stable and post-hospital populations, and studies were subgrouped accordingly for meta-analyses. Studies were also subgrouped according to study design (RCTs or non-randomised controlled studies).

Given probable residual clinical heterogeneity, a random-effects model was deemed most appropriate. The I² statistic (which gives the percentage of the total variability in the data caused by between-study heterogeneity) was reported and commented on where appropriate. Further subgroup analysis (beyond stable and post-hospital populations or study design) was considered, subject to sufficient numbers of studies being available.

Construction of funnel plots to aid assessment of potential publication bias was planned where there were at least 10 studies in a meta-analysis. Sensitivity analyses, that is assessments of the robustness of any meta-analysis conclusions to the inclusion/exclusion of low-quality studies (those at most risk of bias), were planned, subject to sufficient numbers of studies and adequately reported quality criteria. Heterogeneity between studies (in terms of population and intervention characteristics) was explored in order to assess the feasibility of an indirect comparison. Small numbers of studies and/or clinical and methodological heterogeneity precluded the undertaking all of these analyses.

Secondary outcomes

Secondary outcome results (for FEV₁, FVC, 6MWD, PaCO₂, PaO₂) were presented in forest plots to show the overall direction of effect, but were not pooled (see Chapter 1, Decision problem, for choice of which secondary outcomes to analyse). The following factors potentially leading to between-study heterogeneity were explored when considering whether or not to undertake meta-analyses: baseline imbalances, adjusted or unadjusted results, time points presented, type of RCT (parallel or crossover) and adequately presented (or calculable) data.

Study and population characteristics

Patient numbers across the 18 RCTs varied between 13 and 201 (total number of patients 1095). Only three RCTs had > 100 patients (by McEvoy et al. in 2009,74 n = 144; Struik et al. in 2014,75 n = 201; and Köhnlein et al. in 2014,76 n = 195). Length of follow-up varied between 3 and 24 months. There were two crossover trials34,77 (3 months’ follow-up). One RCT was reported as an abstract only;78 this was an interim (3 month) report of the ongoing HOT-HMV (Home Oxygen Therapy versus Home Mechanical Ventillation) trial (see Ongoing studies). One RCT was reported in two publications (by Duiverman et al. in 200879 and 201180), the first79 reporting outcomes after 3 months, during which time usual care in all patients took the form of a rehabilitation programme, and the second80 reporting over a longer follow-up period, with usual care now switched to less-intensive home-based rehabilitation.

Patient numbers across the 10 non-randomised controlled studies varied between 34 and 140 (total number of patients 664). Five were prospective studies, in which a group of patients on NIV and a group of patients receiving only usual care (usually those who refused NIV) were followed up over a period of between 12 and 35 months. Five were retrospective analyses of previously gathered data, with follow-up/analysis periods of 6 months, 12 months and 4, 8 and 10 years (see Study quality on quality assessment for potential selection bias in non-randomised studies).

The average age of patients across RCTs and non-randomised controlled studies varied between 60 and 73 years, and there were typically more men than women (between 41% and 98% men).

Most RCT populations were described as having GOLD stage 3 or 4, had mean FEV₁ % predicted values that were indicative of severe/very severe COPD and/or were described as severe (with no indication of the classification system used). One RCT81 reported no details. There are unlikely to be major differences between RCTs in terms of severity, although the relative proportions of patients with stage 3 and 4 disease are unknown. Most RCTs described populations as being hypercapnic, although the PaCO₂ cut-off points for inclusion varied (e.g. > 6 kPa, > 6.6 kPa). Three RCTs34,81,82 presented mean PaCO₂ levels only, which were suggestive of at least a proportion of patients being hypercapnic. Two RCTs83,84 included normocapnic patients, while one85 stated that the number of hypercapnic patients included was small.

Seven of the non-randomised controlled studies also appeared to include mainly GOLD stage 3 and 4 patients; three studies86–88 provided no details. Eight studies included hypercapnic patients (based on PaCO₂ cut-off points for seven), one86 presented mean PaCO₂ levels only, which were suggestive of at least a proportion of patients being hypercapnic, and a further study87 provided no details.

A distinction has been made between ‘stable’ patient populations and those who commenced NIV after hospitalisation. In the stable populations, authors of the studies have specified that patients should not have been hospitalised within a certain time period. This time period varies between studies (from 4 weeks to 3 months). Not all studies with stable populations have defined a time period but may merely have described patients as stable. For populations described as ‘post hospital’, there was clear evidence in the study report that treatment with NIV commenced after an episode of hospitalisation (because of an exacerbation). There is no information, however, on timings, for example how soon after an exacerbation were patients admitted or how long the period in hospital was before NIV was initiated. This classification has informed the presentation of results (see Clinical effectiveness results: non-invasive ventilation versus usual care and Clinical effectiveness results: non-invasive ventilation versus non-invasive ventilation) and the economic evaluation scenarios (see Clinical effectiveness review discussion). Note that there is limited information on time since last exacerbation (for either population), with the exception of the controlled study by Paone et al. (2014)89 where all patients were enrolled 3 months after discharge from hospital after an exacerbation; they were free from exacerbations for at least 4 weeks and, therefore classified as stable.

Most RCTs included stable populations (13 out of 18), with five75,78,90–92 including a post-hospital population. Four non-randomised controlled studies89,93–95 included stable populations, four86,96–98 included post-hospital populations and there was no description for two. 87,88

Data were sought on patients’ exacerbation history in order to make an assessment of whether any could be described as frequent exacerbators. Most studies provided no information. One RCT92 stated that all patients had had previous exacerbations and three non-randomised controlled studies in stable populations89,93,94 specified at least one previous admission because of severe exacerbation.

Most RCTs (13 out of 18) provided details on assessing patients for obstructive sleep apnoea, for the purpose of ruling out patients with co-existing disorders (overlap syndrome); there were no details for five. 76,78,81,84,92 Only four out of nine89,93–95 non-randomised controlled studies provided clear details.

Details of the study and population characteristics of the RCTs can be found in Table 4 and for the non-randomised studies in Table 5.

Intervention and comparator characteristics

A number of devices were used to administer NIV, reflecting preferences of different countries and changes in devices over time (see Appendix 6). Masks were nasal, oronasal or full-face, sometimes depending on patient choice.

Inspiratory (IPAP) and expiratory pressures (EPAP) were set according to target pressures, target volumes or target blood gases. IPAP settings were compared across studies to ascertain whether or not there were any differences in pressure (see Appendix 6). Pressure was inconsistently reported, for example as mean or median, at the start of the study or at discharge, or described as ‘highest tolerated pressure’, adjusted to patient, or at a level to achieve specific blood gas pressures. Not all studies gave numerical values. Given this lack of consistency and the absence of an agreed cut-off point, it was not possible to dichotomise studies according to high/low pressure (see Clinical effectiveness results: non-invasive ventilation versus non-invasive ventilation for RCTs comparing different NIV settings directly). RCTs with the highest IPAP (mean ≥ 20, where described) in the current set of studies were Duiverman et al. (in 200879 and 201180), Murphy et al. (2011)78 (interim results from the ongoing HOT-HMV study) and Köhnlein et al. (2014). 76 The non-randomised controlled studies with the highest pressures (mean ≥ 20 cm H₂O, where described) were Budweiser et al. (2007)96 and Heinemann et al. (2011). 97

Patients in both treatment arms continued to receive usual care, which was normally standard medical therapy, including LTOT where required, to optimise symptom control (Table 6 shows proportion of patients on LTOT in the RCTs/controlled studies). In 10 studies most (> 90%) or all patients were on LTOT, a smaller or unknown proportion of patients were on LTOT in 11 studies, six studies gave no details and in one study only 4% were on LTOT. 84 It is likely that there were regional variations in what constitutes usual care.

Three studies had what could be considered to be more intensive usual care: the RCT by Duiverman et al. (200879 and 201180) started with patients who were in a 12-week (in-hospital or outpatient) multidisciplinary rehabilitation programme, followed by a long-term home-based rehabilitation programme (one or two sessions per week of physiotherapy at community practice); the patients in the study by Garrod et al. (2000)84 were also in a pulmonary rehabilitation programme for part of the RCT; and the non-randomised controlled study by Clini et al. (1996)94 included as usual care a ‘home supervision programme’ [including physical, occupational and dietary information, a link service between hospital and community health service via telephone contact with general practitioners (GPs) and with patients, monthly physician visits to assess treatment and give further advice and the checking of equipment and decisions on hospitalisation]. It is unclear how additional usual care might influence the effect of NIV.

Study quality

Full quality assessment details can be found in Appendix 2. For the 18 RCTS, risk of bias regarding random sequence generation was low (10 studies) or unclear (eight studies); risk of bias regarding allocation concealment was also low (seven studies) or unclear (11 studies). Blinding of patients was not possible unless a ‘placebo NIV’ was used; this was the case in three studies. Cheung et al. (2010)90 carried out an open-label study which used CPAP in the non-NIV arm and stated that ‘care had been taken to avoid biasing the patients into believing either mode was superior’. Sin et al. (2007)82 also used CPAP as sham therapy, while Gay et al. (1996)100 used the same NIV equipment in the usual-care arm but ‘ventilated’ with lowest EPAP level and had no added IPAP or timed breaths.

There was little information overall on blinding of outcome assessors, with only three RCTs giving some detail (Clini et al. 2002,99 Sin et al. 200782 and Köhnlein et al. 201476). For more objective outcomes, such as survival and hospitalisations, a lack of blinding may be less important than for more subjective assessments such as QoL questionnaires. However, a lack of blinding may also inadvertently lead to performance bias, with a potential for NIV patients to receive more medical attention, which could in turn have an effect on any outcome.

A risk-of-bias rating for incomplete data/handling of missing data was derived as described in the methods (see Chapter 3, Assessment of risk of bias). As noted in the methods section, the cut-off points are arbitrary and a lack of reporting may in some cases be contributing to a high risk-of-bias rating. A high loss to follow-up for outcome assessments which require clinic attendance may to some extent be a result of the nature of severe COPD, with patients finding it difficult to travel. There were nine RCTs34,75–77,79–83,100 with a high risk of bias for incomplete data in the NIV and/or the usual-care arm (for at least one outcome). None of these contributed to the primary outcome meta-analyses. However, it was noted in the relevant results sections where RCTs had a high risk of bias for incomplete data.

Two34,77 of the 15 RCTs had a crossover design. Both appeared to use appropriate statistical methods, although the possibility of a carry-over effect was not explored for one. 77 Both were at high risk of bias for incomplete data, and it in some cases it was unclear during which treatment period patients dropped out.

No sensitivity analysis was performed on the basis of the quality assessment, as (1) a lack of reporting hampered classifying studies according to quality and (2) a lack of reporting is not necessarily an indication of poor study quality. Ideally, a risk-of-bias rating would have been generated for each outcome within each study. Where reported, a distinction has been made for blinding and incomplete outcome data for the different outcomes (see Appendix 2); however, this information was frequently not available. Overall, incomplete data are less likely to be an issue for mortality and hospitalisations/severe exacerbations, while outcome measurements that required patients to attend a clinic are more prone to missing data.

The non-randomised controlled studies were not given a risk-of-bias rating. Of the 10 controlled studies, five89,93–96 had a prospective design and five appeared to be retrospective analyses of data, although this was not always clearly described. One prospective study89 used a matched design, which may result in more similar groups at baseline and thus less biased results. Retrospective studies are more prone to bias, as outcome measurements and usual care cannot be retrospectively standardised for both groups. Most usual-care groups were made up of patients who were eligible for NIV, but could not adhere (e.g. because of mask intolerance) or did not want to continue; however, in three retrospective studies,86,88,97 levels of blood gases determined eligibility for NIV, and one retrospective study98 gave no details. All prospective studies89,93–96 and one of the retrospective studies98 gave details on similarity between NIV and usual-care groups in terms of baseline characteristics, and there appeared to be no major differences. There were no details in one retrospective study,87 and the other three may have differences in blood gas levels as these determined eligibility for NIV. Three studies86,96,97 report that NIV patients had more follow-up visits (at hospital or home) than usual-care patients, which may have impacted on the effectiveness of their usual care.

There were very few details on blinding and frequently incomplete details on losses to follow-up, whether or not intention-to-treat (ITT) analysis was performed and/or how many patients were contributing to results at different time points. Four of the five retrospective studies reported only survival, in which case loss to follow-up may be less of a problem in terms of accessing outcome data.

The potential impact of quality findings are discussed in the individual results sections; however, a formal sensitivity analysis based on quality was not undertaken for non-randomised studies (because of the small number of studies in meta-analyses and difficulties in ascertaining quality cut-off points).

Results from RCTs and non-randomised studies were not pooled because the inherent differences in susceptibility to bias.

Overview of outcomes

Table 7 shows the primary and secondary outcomes (as defined in this report, not by individual study authors) where data have been extracted and analysed. Where studies have reported hospitalisation as a result of an exacerbation, this has been included in the results only once (under hospitalisation) in order to avoid double-counting (e.g. Tsolaki et al. 200895). Where a study has reported severe exacerbations (Cheung et al. 201090), it has been assumed that patients will be hospitalised and this has also been counted as hospitalisation. Where exacerbations have been reported without any indication of severity, they have been presented separately.

There were only very limited data contained within RCTs or non-randomised controlled studies that linked exacerbation, hospitalisation, QoL and survival data (i.e. number of exacerbations leading to hospitalisation, number of exacerbations resulting in death, number of hospitalisations resulting in death or effect of exacerbation and/or hospitalisation on QoL). These outcomes have thus been considered separately. Uncontrolled studies were not searched for linked data.

Primary outcomes were less frequently measured than secondary outcomes (with the exception of adherence to NIV) in the prospective controlled studies, and only a proportion of results could be incorporated into meta-analyses (see Chapter 4, Survival and Hospitalisations individual results sections for more details). Most prospective studies (except two78,96) reported secondary outcomes (one or more of FEV₁, FVC, PaO₂, PaCO₂ or 6MWD). Secondary outcome results from RCTs only are presented in forest plots (see Secondary outcomes). These outcomes were not measured/reported by the retrospective controlled studies.

Survival

Ten RCTs reported survival, seven in stable populations and three in post-hospital populations. All were meta-analysed or presented in a forest plot. All 10 non-randomised controlled studies reported survival, and all were presented in one or more forest plots. Follow-up times for all but one of the RCTs was between 6 and 24 months only; where follow-up was beyond 24 months (in one RCT and four controlled studies), the results have been presented in a separate plot. Relative risks (RRs) and hazard ratios (HRs) were taken directly from the published article where possible. Otherwise they were calculated as described in Chapter 3, Analysis (RR for all populations, HR for stable populations only). A footnote indicates where RR or HR was calculated from event rates reported in the published article.

Figure 3 shows the RR for mortality for all studies, subgrouped by RCTs/non-randomised studies and stable/post-hospital populations. Pooling was undertaken only within subgroups. There was no difference in mortality between stable populations treated or not treated with NIV, based on seven RCTs and four controlled studies. One study88 which reported no details on the population has been categorised separately; this also found no difference. There was also no difference in mortality for the post-hospital population based on three RCTs, but a statistically significant difference favouring NIV based on four non-randomised studies in post-hospital populations.

FIGURE 3.

Mortality (RR). CI, confidence interval. a, Calculated by authors of this report; b, controlled study with matching.

Between 57% and 100% of patients within the stable populations were on LTOT (unclear for NIV arm in Zhou et al. 200881); this was also the case for the three post-hospital studies. 75,92,98 One post-hospital study (Budweiser et al. 200796) included around 50% patients on LTOT at the start of the study; however, at discharge most were on LTOT. A subgroup analysis for patients with or without LTOT found no difference. Two studies86,97 gave no details on LTOT.

Figure 4 shows the HRs, again pooled by subgroup (RCT/controlled study, stable/post hospital). Unadjusted HRs showed no difference between treatment arms for the stable populations, while the adjusted HR from the McEvoy et al. 2009 RCT74 was significantly in favour of NIV (adjusted for those variables found to be confounders based on changing the HR by 10% or more when added to the model: PaCO₂, PaO₂ and SGRQ total score). It should be noted that the upper confidence limit for the Köhnlein et al. RCT76 differs slightly from that in the published article, as the standard error (SE) had to be approximated from the data given.

FIGURE 4.

Mortality (HR). CI, confidence interval. a, Calculated by authors of this report. b, Controlled study with matching.

The adjusted and unadjusted HRs reported in two controlled studies in post-hospital populations were significantly in favour of NIV (adjusted for LTOT, haemoglobin, age and body mass index96 and for age, PaO₂, haemoglobin and haematocrit at discharge97). However, these are observational studies and there is no way of knowing whether or not the estimated HRs are biased because of unmeasured or inappropriate adjustment for confounders.

Of the above studies, those with the highest IPAP were by Budweiser et al. (2007)96 and Heinemann (2011),97 both non-randomised studies. The fact that these studies showed a significant difference in favour of NIV may be a result of the higher pressure and/or of a greater benefit from NIV in a post-hospital population. Patients in both studies had more frequent evaluations in the NIV arm, and there may be additional confounders which have not been adjusted for.

Figure 5 shows mortality at longer follow-up times (> 2 years). Data are not pooled as multiple time points are presented. One RCT in a post-hospital population found no significant difference. 75 Three controlled studies, two with no details on the population and one in a stable population,93 also found no difference. The one post-hospital study86 showed a statistically significant benefit from NIV at 2 and 5 years, with less of a benefit at 10 years (although this was still statistically significant). It is conceivable that any potential benefit from NIV would not be sustained for the duration of a patient’s life given the progressive worsening of COPD. There were no details on IPAP settings for this study. The schedule of NIV was also unusual in this study and is not consistent with all other studies, as it was used intermittently for 15 minutes every hour, up to a minimum of 4 hours. For all studies an assumption has been made that the survival rates are based on no loss to follow-up; patient numbers have been calculated from percentages and rounding errors may have occurred. Given that the population details were unclear in two studies, it was difficult to draw any conclusions. It could be speculated that if the study had specifically included a post-hospital population this may have been highlighted.

FIGURE 5.

Mortality (RR) longer follow-up. CI, confidence interval. a, Calculated by authors of this report.

Overall, the results on mortality (RR, unadjusted HR) suggest that, on average, there is no benefit from starting NIV in a stable COPD population, and this is consistent for RCTs and non-randomised controlled studies (n = 770 in total). This is based on a fairly limited follow-up period for mortality (up to 24 months for RCTs), and it is possible that small differences might not manifest themselves within this time frame. One exception is the adjusted HR (McEvoy et al. 200974), which shows a statistically significant benefit for NIV.

The results from three RCTs in a post-hospital population also show that, on average, there is no benefit to home NIV. One further RCT (Xiang et al. 2007,92 n = 40; not represented in the forest plots) also found no statistically significant difference. This is in contrast to the non-randomised controlled studies in this population, which show a statistically significant pooled result favouring NIV. Follow-up in the RCTs was 1290 and 2492 months, arguably not long enough to measure mortality as an outcome. The sample size was small in two RCTs (40 patients in each) and larger in one RCT (Struik et al. 2014,75 n = 201).

Further subgroup analyses (e.g. by level of hypercapnia or history of exacerbations) were not possible given the small number of studies and the inconsistent reporting of the relevant variables. Based on the inclusion criteria (CO₂ threshold) and description of populations, most studies appeared to be on hypercapnic patients, with the exception of that by Casanova et al. 2000,85 who stated that ‘The number of hypercapnic patients in our series was small’. Performing subgroup analysis based on reported mean baseline CO₂ values would have meant dichotomising trials based on an arbitrary threshold, and this was not considered appropriate. However, in a separate analysis, CO₂ levels at baseline and change in CO₂ levels were plotted against mortality, in order to determine if baseline CO₂ can predict response to NIV and whether or not the effect of NIV on CO₂ correlates with the effect on mortality (see Figure 42 and Figure 44, Appendix 7). There was no discernible relationship either between CO₂ levels at baseline and mortality or between change in CO₂ levels and mortality. This was an exploratory post-hoc analysis and subject to a number of limitations; as such the findings should be considered to be speculative only and should be interpreted with caution (see Appendix 7 for caveats).

Sensitivity analyses by study quality were also not performed, again because of the small number of studies, and the lack of reporting of details on quality. Only one of the RCTs80 was flagged as having a high risk of bias regarding incomplete data. Only one RCT90 used a form of sham NIV, so patients were not blinded in most studies. There may also be a risk of performance bias if the use of NIV is associated with additional check-ups or follow-up visits. Assessment of potential publication or other bias through funnel plots was not undertaken given that there was a maximum of seven studies in any given meta-analysis.

Non-randomised studies are more prone to overestimating effect size; this does not seem to be the case here for stable populations, as the results of RCTs and non-randomised controlled studies are consistent. The non-randomised studies in the post-hospital population do show a benefit of NIV, in contrast to the pooled result from RCTs in this population (which shows no difference).

The potential for selection bias was explored for the non-randomised studies. NIV and usual-care arms appeared to be similar for the stable populations (see Appendix 2 for full details), whereas Laier-Groeneveld and Criee (1995)88 (population unclear) reported that patients in the usual-care group were normocapnic and those in the NIV group were hypercapnic. For the post-hospital populations, baseline characteristics also appeared to be similar in two studies,96,98 although one98 gave no detail on how groups were selected. In the other two studies,86,97 blood gas levels determined the eligibility for NIV; that is, only patients with poorer blood gas levels/lower level of hypercapnia received NIV.

Differences in usual care were also explored. In the studies by Budweiser et al. (2007),96 Heinemann et al. (2011)97 and Milane and Jonquet (1985)86 (all post-hospital populations), patients in the NIV groups received more frequent hospital evaluations or home visits, which may have led to better usual care. These studies found a statistically significant benefit for NIV.

Uncontrolled observational studies were explored in order to determine whether there were additional useful data from larger and/or longer-term studies (see Appendix 8 for further details). There were two larger prospective studies102,103 following stable patients on NIV. Survival rates in one103 of these were similar to those reported in the NIV arm of Laier-Groeneveld and Criee (1995),88 while the other102 reported slightly lower survival rates. An analysis of patient survival depending on haemoglobin level (Kollert et al. 2013104) reported survival rates for normocythaemic patients (n = 207) of 72%, 50%, 47% and 18% respectively at 2, 4, 5 and 10 years (approximate, estimated from graph); these were similar at 2 years and slightly lower at 4 years compared with Laier-Groeneveld and Criee (1995),88 although a comparison is difficult as nothing is known about the haemoglobin status in the controlled studies included in this report. One small study105 followed both stable (n = 16) and post-hospital (n = 31) patients on NIV. This found significantly poorer median survival in post-hospital patients compared with stable patients, which is consistent with the population differences observed in the controlled studies.

Hospitalisations

Hospitalisation data were reported in a number of ways, for example the number of hospitalisations (general or ICU) per patient per year, mean number of days in hospital, proportion of patients affected by hospitalisation, time to first readmission, etc. Only the first two outcomes are represented in Figures 6 and 7, based on eight RCTs (five76,80,81,99,101 in stable and two90,92 in post-hospital populations) and three controlled studies93–95 (all stable populations). Severe exacerbations have been assumed to lead to hospitalisation and have been included in the hospitalisation analysis. The hospital admissions and days in hospital were all presented per patient per year, except for the Clini et al. (1996)94 study, which reported hospitalisation rates over a period of 18 months; the results have been adjusted to 1 year (which assumes a constant rate). Additional hospitalisation data not represented in the forest plots are shown in Table 8.

For stable populations, there appeared to be a non-significant trend towards fewer hospital admissions with NIV (based on five RCTs) and ICU admissions (based on one RCT and two controlled studies); no difference in hospital admissions was found based on two controlled studies (Figure 6).

FIGURE 6.

Hospital admissions per patient per year. CI, confidence interval; WMD, weighted mean difference. a, Calculated by authors if this report; b, individual mean differences (95% confidence interval) presented for this outcome.

Note that the forest plot does not include results from all RCTs; McEvoy et al. (2009),74 who carried out a fairly large RCT (n = 144), and Casanova et al. (2000),85 who studied 52 patients, reported very similar hospitalisation rates74 and proportion of patients affected by hospitalisations85 in the NIV and usual-care groups. Had it been possible to incorporate these results, the overall pooled estimate may have shifted more towards an equivocal effect (see Table 8 for additional results). It should be noted that the Köhnlein et al. 76 study in 2014 is given a relatively small weighting in the meta-analysis, despite it being a larger study (n = 195); this is because the standard deviation (SD) (and thus the SE) was much higher. It could be argued that hospital admissions data are by their nature skewed (with some patients having repeated admissions and one admission making a further one more likely) and that the median may be a better metric to compare groups. However, all but two of the studies in the meta-analysis reported the mean in their publications; the authors of two studies75,80 reported the median but provided the mean (SD) on request.

For the post-hospital population, the results from the three RCTs give disparate results (significant difference in favour of NIV92, non-significant difference in favour of NIV90 and non-significant difference on favour of usual care75). The results have therefore not been pooled. Note that the (larger) study by Struik et al. 75 in 2014 has a similar weight in the meta-analysis compared with the smaller studies; this is because the SD (and thus the SE) was much higher. The study by Cheung et al. (2010)90 randomised patients to NIV or CPAP and, while this was an open-label study, the authors stressed that ‘care had been taken to avoid biasing the patients into believing either mode was superior’. There was no sham NIV in the studies by Xiang et al. (2007)92 and Struik et al. (2014),75 and arguably the results from the Cheung et al. 90 study in 2010 could be considered more robust. However, it is likely that there are additional clinical or methodological differences between the studies.

Days in hospital (per patient per year) were reported, only for stable populations, in one RCT and three controlled studies (Figure 7). There was no significant difference, although a very slight trend towards benefit with NIV. Again, this is not based on the totality of the evidence and firm conclusions cannot be drawn.

FIGURE 7.

Days in hospital per patient per year. CI, confidence interval; WMD, weighted mean difference

Additional data on hospitalisation are reported in Table 8 (note that some studies have used more than one outcome measure and are represented in both Figure 6 and Table 8). There were no significant differences between NIV and usual care based on three RCTs (and one small crossover RCT) and one controlled study (stable population), with the exception of proportion of patients affected by hospitalisations at 3 months in one study (Casanova et al. 200085). One matched controlled study89 found a significant difference in favour of NIV at 24 months.

Three RCTs in post-hospital populations also found no significant differences, although there appeared to be a trend for a longer time to first readmission due to any COPD exacerbation with NIV (Cheung et al. 201090).

Based on the meta-analyses, there appeared to be an overall trend towards fewer hospitalisations with NIV (for a stable population); however, this is not based on the totality of the evidence. For a post-hospital population, the results are quite disparate and likely to be because of clinical and/or methodological heterogeneity or, particularly in the case of the two smaller studies (n = 40 and n = 47), chance. Arguably, the Struik et al. (2014)75 study, which suggests no benefit from NIV in this population, could be given more consideration, as it is based on a much larger sample (n = 201). Looking at all the evidence, no firm conclusions could be drawn as to whether or not stable or post-hospital patients are likely to benefit from NIV in terms of fewer hospital admissions or days in hospital, although there may be patients with specific characteristics who would benefit.

None of the five included RCTs (stable population) had a sham NIV arm, so lack of blinding may be a source of bias. Sensitivity analyses around study quality were not undertaken because of the small number of trials and the lack of reporting of some quality criteria. Similarly, funnel plots were not constructed to assess potential for publication or other bias, as there were a maximum of five studies in any given meta-analysis.

Further subgroup analyses (e.g. by level of hypercapnia, history of exacerbations, proportion on LTOT or IPAP) were not possible given the small number of studies and the inconsistent reporting of the relevant variables. As with mortality data, baseline CO₂ levels and change in CO₂ levels were plotted against mean difference in hospital admissions in order to determine whether baseline CO₂ can predict response to NIV and whether or not the effect of NIV on CO₂ correlates with the effect on admissions (see Figure 43 and Figure 45, Appendix 7). It was unclear whether there was a trend towards an association between higher mean CO₂ levels at baseline and a greater mean difference in hospital admissions, but an apparent trend was observed for change in CO₂ levels and admissions. Such a trend would indicate that a greater effect in terms of reducing CO₂ levels correlates with a greater reduction in hospital admissions. However, as the analysis is using aggregate data for change in CO₂ and also for mean difference in hospital admissions, a causal association cannot be inferred, even if there is potential biological plausibility. Further, this was an exploratory post-hoc analysis and subject to a number of limitations; as such, the findings should be considered speculative only and should be interpreted with caution (see Appendix 7 for caveats).

There were three small prospective uncontrolled observational studies, two106,107 in stable populations (n = 11 and n = 35) and one108 in a post-hospital population (n = 27) (see Appendix 8 for further details). The study in a post-hospital population did not report before and after (NIV) data and so did not add useful information. The studies in stable populations found a reduction in admissions/days in hospital over time with NIV (one pre–post NIV106 and one after 1 year compared with after 2 years with NIV107). This needs to be interpreted in the context of the small number of patients and the lack of a usual-care group, but might be consistent with the trend observed in some of the controlled studies.

Exacerbations

Severe exacerbations (assumed to be exacerbations leading to hospitalisation) are described previously (see Chapter 4, Hospitalisations). This section has attempted to identify additional exacerbations (mild or moderate) which do not lead to hospitalisation but could potentially incur a cost and/or have an impact on QoL (Table 9). However, this was hampered by a lack of reporting of severity. In some cases, exacerbations were not a predefined outcome but instead were listed as an adverse event (Garrod et al. 2000,84 Meecham-Jones et al. 199577). As it was unclear whether or not all exacerbations were reported (rather than only those in the context of not completing final assessment84 or in the context of cause of death77), the data from these two studies has not been further considered.

Pooling of exacerbations data was not possible, as a number of different outcome statistics were used; these could not be converted to make them more consistent.

There was little information on the frequency of exacerbations of differing severity (mild, moderate or severe). Cheung et al. (2010)90 noted exacerbations without acute HRF, which may be less severe, while Tsolaki et al. (2008)95 made a distinction between all exacerbations and those leading to hospitalisation. Struik et al. (2014)75 reported exacerbations occurring at home.

Overall there were no significant differences in exacerbations between NIV and usual-care arms, based on three RCTs and one controlled study in stable populations and two RCTs in a post-hospital population. One exception was the study by Zhou et al. (2008)81 (n = 36, stable population), which reported significantly fewer exacerbations with NIV (at 12 months). Subgroup analyses were not possible. IPAP was highest in the study by Duiverman et al. (2011)79 (mean 23 at start of study), and between 12 and 20 in the other studies (where reported). The study by Duiverman et al. (2011)79 also had a more intensive usual-care arm. The study by Bhatt et al. (2013)83 differed from the others in that it included normocapnic patients. Around half of80,90 or most85 patients were on LTOT, with no details in two studies. 83,95

Given the sparsity of information, the uncontrolled studies were also considered. Only three studies (no restriction on sample size or length of follow-up) reported exacerbations. Neither the prospective uncontrolled study by Tsolaki107 in 2011 nor the retrospective study by Windisch109 in 2009, both in stable populations, made a distinction between different severities of exacerbation. The report of a further small (n = 20) Japanese study64 was not translated. It is possible that reporting of exacerbations was included within hospitalisation and/or adverse outcome data; however, this would have necessitated reading the full texts of all the uncontrolled studies, which was not undertaken.

Quality of life

Seven RCTs and one prospective controlled study (Tsolaki et al. 200895), all with stable populations, reported QoL using a variety of instruments (SF-36, SGRQ, SRI, the Chronic Respiratory Disease Questionnaire, the MRF-28 and the Profile of Mood States). Note that for the SGRQ, the MRF-28 and the Profile of Mood States, a lower score is indicative of better QoL (or more stable mood profile for Profile of Mood States). For the other instruments, a higher score is indicative of better QoL.

There was heterogeneity regarding the number of QoL instruments used (between one and three), time points of reporting QoL (between 3 and 24 months), and in the way results were reported, for example for subscales only rather than the total score or as a mean difference between final scores rather than a change score or adjusted score. Some studies did not report numerical data but only commented on statistical significance or only presented results on a graph. Further, in order to pool, an assumption would have had to have been made that all instruments were measuring QoL in the same way and were equally valid. For these reasons it was not considered feasible to pool any QoL data.

Table 10 presents the findings. Some studies used several QoL instruments, thus contributing to a greater proportion of the findings (e.g. Duiverman in 2008 and 2011,79,80 which also had a more intensive usual-care arm compared with the usual care employed in most studies). Note that, for the study by Köhnlein et al. (2014),76 results were available only for a small subgroup of patients.

The overall results were suggestive of trend towards better QoL with NIV compared with usual care, with some results statistically significant (in favour of NIV). This is not consistent across all studies, however, with one of the larger RCTs (McEvoy et al. 200974) finding statistically significant results in favour of usual care, albeit for only some subscales. The one non-randomised study may have been prone to greater bias than the RCTs; however, the authors stated that there were no statistically significant differences between groups for baseline characteristics (thus making selection bias less probable).

The 2013 Cochrane review110 pooled 12-month SGRQ results for McEvoy et al. 200974 and Clini et al. 200299 based on individual patient data (IPD) (which were not available for this report). They found a very small, and not statistically significant, effect in favour of the usual-care group [mean difference of 0.9, 95% confidence interval (CI) –19.21 to 21.01]. This result is based on one QoL instrument only. See also Chapter 5, Review of systematic reviews.

Only one RCT (Struik et al. 201475) in a post-hospital population reported QoL; there were no statistically significant differences between groups at 12 months, based on four different measurement scales.

None of the studies reporting QoL had a sham NIV arm, so there may be a risk of patients receiving NIV being more optimistic in their assessment of QoL. However, NIV is also known to have an adverse impact on QoL in the short term (while patients are adapting).

Given the paucity of QoL data in a post-hospital population, the uncontrolled studies were explored (only studies with at least 10 patients considered, using one of the QoL instruments used by the RCTs/controlled studies). There was only one small prospective uncontrolled study (Skobel et al. 2011108) which reported QoL using a German version of the SGRQ at 3 months in 27 patients. This study found a significant improvement in the symptom and impact domains, as well as in overall score compared with the baseline; after further follow-up the symptom score showed a deterioration (while the scores for the other domains remained stable). There are too few data to make an assessment of how QoL changes might differ in stable and post-hospital populations.

There were no retrospective uncontrolled studies conducted in post-hospital populations using one of more of the relevant QoL instruments.

Adherence and adverse events

Table 11 shows details of the recommended schedule of NIV, any support provided for adapting to NIV, the extent to which patients adhered to NIV and any adverse events associated with NIV, including where these led to treatment discontinuation. Length of hours that NIV is used for per night and the percentage of nights NIV is used for may affect the potential benefit. Overall, there was a lack of consistency and/or detail in terms of how these parameters were reported across studies. Generally there was less detail in the retrospective studies.

Most studies reported some detail on how patients were accustomed to NIV; patients usually had a run-in period in hospital [between 2 and 6 days where stated, with the exception of Clini et al. (1998)93 with 15 days], and some studies gave details of follow-up support in the home. NIV was generally recommended for nocturnal use, with the exception of Milane and Jonquet (1985),86 who specified a schedule of 15 minutes per hour when awake, up to a minimum of 4 hours per day, and Zhou et al. (2008),81 who also specified a more intermittent schedule (three sessions over a 24-hour period). Some studies recommended a minimum number of hours of use (between 5 and 9); one study74 defined consistent use as an average of > 4 hour per night.

Most of the prospective studies reported the average number of hours during which NIV was used each day (assessed by ventilator recordings), but only three76,83,84 gave details on the percentage of patients using NIV for a minimum number of hours on average. Average use may not be informative regarding the proportion of patients likely to benefit from adequate use of NIV. The mean use varied between 3.1 and 9.2 hours per night, although it was not always stated if this referred to completers only. Variations may be a reflection of recommendations given by clinical staff, disease severity or patient comfort (e.g. different NIV pressures, ability to sleep with equipment). No retrospective controlled studies had details on average NIV use, as this information would likely not have been captured.

There was variation in how studies reported adverse events (e.g. all associated with NIV or only those associated with discontinuation), hampering comparisons across studies. Reasons for discontinuation due to NIV included high pressure, inability to sleep, disturbance of partner, mask intolerance/claustrophobia, perceived lack of effect or general discomfort or intolerance. There was one case of suspected barotrauma. 92 Rates of discontinuation from NIV varied between 5% and 43%; these withdrawal rates were not all because of adverse events but may have included other reasons (e.g. other illness). Differing discontinuation rates may be a reflection of length of follow-up, although most discontinuation would be expected soon after starting NIV. Other adverse events, not necessarily leading to discontinuation, included skin lesion/inflammation, dry mouth/throat/eyes (humidification was provided in some cases), rhinorrhoea, gastric distension and anxiety.

Secondary outcomes

Results for key secondary outcomes are presented below. All results are based on RCTs only; non-randomised controlled studies were not considered. Results are presented in forest plots in order to show the overall direction of effect; however, pooling was not undertaken owing to baseline imbalances between arms in many of the RCTs and a lack of reporting of adjusted results. Additional reasons for not pooling were differences between studies in terms of time points presented, the type of RCT (e.g. parallel or crossover) and, in some cases, uncertainty around patient numbers or exact results (where estimated from graphs or units were converted; see also Appendix 3 for details on calculation/assumptions made). Some studies are represented more than once (results at different time points or reported using different metrics). All extracted (or calculated) data relating to secondary outcomes are presented in Appendix 9.

Only one study32,79 presented adjusted results using appropriate methods (linear regression analysis). Some studies presented change scores; for most studies mean difference between post-treatment scores was calculated for this report. Studies were separated in forest plots depending on the method used for calculating a mean difference (adjusted, change score or mean difference based on final scores). The Köhnlein et al. 76 RCT presented all results for secondary outcomes as percentage change only, therefore results could not be incorporated into forest plots. Findings have been described narratively for relevant outcomes.

Forced expiratory volume in 1 second

Five RCTs,34,74,83,85,99 all in a stable population, reported FEV₁ (% predicted), at time points between 3 and 24 months (Figure 8). Two studies34,99 reported a change score; for the other three a mean difference was calculated from aggregate final scores. There were no significant differences in any studies and no overall trend was observed.

FIGURE 8.

The FEV₁ % predicted. a, Calculated by authors of this report.

Six RCTs32,34,77,79,84,100,101 in stable populations and two75,92 in a post-hospital population reported FEV₁ (l), at time points between 3 and 24 months. All but one,84 which did not report sufficient data, have been represented in a forest plot (Figure 9). There may be a very slight trend across studies favouring NIV, with three statistically significant results, all of which show a clinically important change (of at least 120 ml). However, two of these were from the same study79,80 in a stable population; this study had more intensive usual care in both arms than in the other studies. The other significant result was in a post-hospital population92 (based on post-treatment scores). Overall, there is a lack of consistency in terms of direction of effect, for both populations.

FIGURE 9.

The FEV₁ (l). a, Calculated by the authors of this report.

Forced expiratory volume in 1 second was measured in two further studies; however, in one82 no post-treatment data were presented (it was stated that there were no significant changes in either arm), and in the other91 the data presented for SDs appeared to be incorrect. The RCT by Köhnlein et al. 76 measured FEV₁ and found a statistically significant difference in favour of NIV.

Forced vital capacity

Two studies in stable populations reported FVC (% predicted; Figure 10). Both results were calculated from post-treatment scores. Neither was statistically significant and the direction of effect is not consistent. Note that the population on the study by Bhatt et al. 201383 was normocapnic and thus not representative of the majority of NIV patients across studies.

FIGURE 10.

The FVC (% predicted). a, Calculated by authors of this report.

Five studies reported FVC (l), which in four cases is presented in a forest plot (Figure 11; one study84 did not report sufficient data to be included). Two studies reported a change score75,77 and two a mean difference calculated from post-treatment scores. Only one study (post-hospital population)92 found a significant difference in favour of NIV. Overall there were few data, and no consistency in terms of direction of effect. The study by Köhnlein et al. 76 reported data for FVC (unclear if % predicted or litres) and found no significant difference.

FIGURE 11.

The FVC (l). a, Calculated by authors of this report.

Partial pressure of carbon dioxide in the arterial blood

Partial pressure of carbon dioxide in the arterial blood was measured in 16 studies, of which 13 are represented in Figure 12. There was an overall trend favouring NIV for reduction in PaCO₂, with seven statistically significant results, one of which was an adjusted result. Mean difference in favour of NIV was between 0.4 and 1.60 (the latter in the post-hospital population). A reduction of 1 kPa might make a clinical difference, but this also depends on the individual patient and their starting value. The proportion of patients within the individual studies achieving a (for them) meaningful change in PaCO₂ is unknown, as is the proportion reaching a threshold values of what might be considered to be ‘normal’ (around 6.5 kPa).

FIGURE 12.

The PaCO₂. a, Calculated by authors of this report; b, measurement performed regardless of oxygen use; c, measurements both on room air or both on oxygen at the same flow rate.

The three studies not represented found a statistically significant difference in favour of NIV,76 found no significant difference82 or appeared to have errors in the data. 111

Partial pressure of oxygen in the arterial blood

Nine RCTs34,77,79–81,83–85,99,100 in a stable population and two75,92 in a post-hospital population reported PaO₂ and could be included in a forest plot (Figure 13). There was a clear trend in favour of NIV for the stable population, with some results statistically significant. There were fewer data for a post-hospital population and no consistent trend. The difference in change for the statistically significant results was between 0.5 and 1.0. Again, a 1 kPa difference might be important but the same caveats as for PaCO₂ apply. Results from only one study79,80 with a stable population were adjusted for baseline (this study had more intensive usual care in both arms). Three further studies (in a stable population) measured PaO₂ but were not included: two74,76 found no significant difference and the other91 appeared to have errors in the reported SD.

FIGURE 13.

The PaO₂. a, Calculated by authors of this report; b, measurement performed regardless of oxygen use; c, measurements both on room air or both on oxygen at the same flow rate.

Six-minute walking distance

Eight RCTs32,77,79,82,83,91,99–101 in a stable population and one92 in a post-hospital population reported 6MWD at baseline and at follow-up times between 3 and 24 months. There was one statistically significant result for the stable population (in favour of NIV, adjusted result), but, overall, the results for the stable population are inconsistent in terms of direction of effect, which may be a reflection of differences in baseline values. The one post-hospital study shows a statistically significant result in favour of NIV (improvement of 86 m), but this is not based on an adjusted result. One study (stable population)77 reported medians (and ranges) only and is not represented in the forest plot (Figure 14). A further study (stable population),76 also not represented, found a statistically significant difference in favour of NIV.

FIGURE 14.

The 6MWD. a, Calculated by the authors of this report.

Discontinuation studies

Two RCTs met the inclusion criteria of this report but addressed a question that was beyond the remit of this report, namely to look at the effect of discontinuing NIV. Both RCTs (Funk et al. 201171 and Oscroft et al. 201072) included patients on NIV who, after a period of withdrawal, were randomised to continue withdrawal or restart NIV. The results have not been further explored here.

Summary clinical effectiveness (non-invasive ventilation versus usual care)

Intervention and comparator characteristics

The three crossover RCTs compared different ventilator settings (Table 15). Dreher et al. 201032 compared high-pressure (assist/control mode) and low-pressure NIV (pressure support mode). This is similar to the comparison in the study by Oscroft et al. 2010,112 which also compares a high-pressure mode with a low-pressure mode.

The study by Murphy et al. (2012)113 compares two high-pressure modes, with high and low back-up rates, the hypothesis being that high inspiratory pressures with low back-up rate would result in levels of ventilator usage equivalent to those of high inspiratory pressures with high back-up rate.

Use of LTOT in the three studies was 100%,32 67%112 or unknown113 (no details provided).

Quality assessment

Full quality assessment details can be found in Appendix 2. Risk of bias for the generation of the randomisation sequence was unclear for all three RCTs; two112,113 provided details of allocation concealment. One study112 stated that trial subjects and technicians were blinded to the ventilation mode, one113 was described as single blind and one33 was open label. Given that all patients received an active treatment, lack of blinding may have been less important than in the studies comparing NIV with usual care only.

Risk of bias from incomplete outcome data was rated as high for two studies. In the study by Murphy et al. (2012),113 42% (5 out of 12) of patients withdrew, four of those during the high-pressure ventilation period; however, it was reported that there were no differences between completers and withdrawers, with the exception of FVC. The other study32 had a loss to follow-up of 23% (4 out of 17) and no details on differences in characteristics between completers and dropouts. Dropouts did not appear to be included in the analysis.

All were crossover trials and appeared to use appropriate forms of analysis; however, only one study32 performed a period-effect test for carryover effects.

Primary outcomes

Given the short duration of the head-to-head trials, the only primary outcome reported was QoL. The studies by Dreher et al. (2010)32 and Oscroft et al. (2010)112 comparing different pressures found no differences in total SRI score32 or SF-36 and SGRQ scores,112 although there was a trend for better QoL on the SGRQ with volume-assured NIV. The first32 of these two studies was rated as having a high risk of bias regarding incomplete outcome data. Neither of these studies was designed to look at the main outcomes of interest for this report as a primary outcome (their main outcomes as defined by the authors of the study were blood gases).

Murphy et al. (2012),113 which compared different breathing frequencies, also found no significant differences in total SRI scores, although there was a statistically significant difference for the respiratory symptom domain in favour of high-pressure ventilation (pressure support ventilation). The risk of bias for incompleteness of data was rated as high. It should be noted that the main aim of this study was to compare ventilator adherence rates. No conclusions can be drawn from these small, short-term trials regarding the effect on QoL of different ventilator settings.

Given the short follow-up times, small numbers of patients, differences in instruments and the inconclusive results from the larger (parallel) trials, it was not feasible to assess the consistency of results across trial design.

Adherence and adverse events

Table 16 gives details of the recommended schedule of NIV, any support provided for familiarisation, the extent to which patients adhered to NIV, and any adverse events associated with NIV, including where these led to treatment discontinuation.

In the study by Oscroft et al. (2010)112 patients were already established on NIV, and in Dreher et al. (2010)32 there was an indication of the number of days patients had needed for initiation of NIV. There were no details in Murphy et al. (2010). 113 All three trials stated that NIV use was nocturnal but there were no details on minimum number of hours recommended. Data were recorded by the ventilators. Treatment compliance was similar between study arms in two studies,112,113 and was higher in the high-intensity arm in one. 32

Adverse events from NIV leading to dropouts included intolerance and claustrophobia; no conclusion could be drawn from this small data set on whether or not (specific) adverse events and/or discontinuations are more common with certain ventilator settings.

Secondary outcomes

The same secondary outcomes were considered for the head-to-head trials (FEV₁, FVC, PaCO₂, PaO₂ and 6MWD). Table 17 shows the outcomes reported by the three trials and the main findings. Overall, there appear to be few significant differences in outcomes, although this may be in part a reflection of the small sample sizes and short follow-up times. There may be a slight trend for a benefit from high-pressure versus low-pressure NIV, with one statistically significant result (reduction in PaCO₂) shown in Dreher et al. (2010). 32 Note that not all outcomes measured in the trials have been reported here.

Summary clinical effectiveness (non-invasive ventilation versus non-invasive ventilation)

• Three small crossover trials with stable populations were identified, two32,112 comparing higher versus lower pressure, and one113 comparing similarly high pressures, with settings differing in back-up rates.

• All trials were short term (6–8 weeks) and did not have the outcomes of primary interest for this report as their primary aim. Dropout rates were high in two32,113 of the studies.

• Treatment compliance was similar between study arms in two studies112,113 and was higher in the high-intensity arm in one. 32

• Quality of life was the only primary outcome measure (based variously on SRI, SF-36 and SGRQ). No firm conclusions could be drawn from the limited data on whether certain settings are more beneficial.

• There was one statistically significant result (Dreher et al. 201032) in terms of PaCO₂ reduction with higher-pressure NIV.

Main findings

Eighteen RCTs and 10 controlled studies were identified that compared domiciliary NIV with usual care. Despite this fairly high number of studies, sample sizes were generally small (between 13 and 201 patients) and not all reported on the primary outcomes relevant to this report. The main findings of studies comparing NIV to usual care, for the two relevant COPD populations, are described below.

Strengths and limitations of the available data

A number of RCTs of reasonably good methodological quality were available, particularly for the stable population. However, there were several limitations, particularly concerning the outcomes relevant to the economic model. Hospital admissions data were not reported consistently across trials, so not all available evidence contributed to the pooled estimate for the stable population; further, admissions data may be skewed and the mean (SD) may not be the most appropriate metric to use, although it was frequently reported. QoL was inconsistently measured in studies of stable patients, and in only one study of post-hospital patients. As a result, the impact of domiciliary NIV on QoL remains uncertain for both populations. A trend of improved QoL needs to be confirmed in stable patients. While a reduction in hospital admissions will likely have a temporary impact on QoL, the effect of NIV on daily living (when COPD is stable) remains uncertain. Further research on which instruments best capture QoL in this patient group and whether scores are translatable between different instruments is necessary. Qualitative research may help to inform which aspects of QoL should be reflected in assessment tools. The timing of QoL measurements is important. Patients may experience a decrease in QoL early on (while they are adapting to the equipment) but an increase at later stages, provided the treatment is effective. A potential survivor effect also needs to be taken into account; QoL may eventually decrease in the later stages of COPD, but this could be in patients who might not have survived without NIV.

For secondary outcomes, most studies did not present adjusted results, despite there being baseline imbalances. Only one study79,80 used an appropriate method of adjusting. Baseline imbalance should be accounted for using an analysis of covariance model. 131 In the absence of this, analyses that use the summary results to adjust for baseline imbalance will not account for the within-study correlations between these repeated measures. More sophisticated analyses may be possible; for example, external evidence could be used to provide an informative prior for the within-study change variance, but this was outside the scope of the current report.

Other limitations in the available data included poor description of when and where NIV was initiated; a lack of detail to allow for a complete judgement on study quality; a lack of reporting of time-to-event data and HRs (even for survival data); inconsistent reporting of adverse events; and a lack of data explicitly linking the number of exacerbations to subsequent hospitalisations and survival. This last point has potential implications for double-counting data, as these outcomes are not independent of each other.

Strengths and limitations of this report

A comprehensive search strategy (including citation checking) meant that this report identified more relevant studies than previous systematic reviews (even taking into account different search periods).

The fact that no language restrictions were applied meant that 6 out of 31 (19%) of the included RCTs or controlled studies were non-English language, constituting a substantial proportion of the overall evidence base.

This is the first systematic review to attempt to account for differing baseline risks of exacerbation/hospitalisation by categorising populations into stable and post hospital. This is also the first systematic review to both focus on patient-related outcomes (survival, hospitalisation, exacerbations and QoL) and incorporate data from non-randomised studies in separate analyses. Further, by calculating summary measures from raw data or converting data, the number of results that could be presented in forest plots was maximised.

In contrast to some previous systematic reviews, secondary outcome data (lung function, blood gases and 6MWD) were not pooled because of a lack of reporting of results which were appropriately adjusted for baseline. A failure to do this can result in misleading results (e.g. where the usual-care arm shows a higher final result but the improvement has been greater in the NIV arm). Instead, results were presented in forest plots with the method of calculating the effect estimate highlighted, so that findings could be put into the context of the robustness of the methods.

Heterogeneity in the way some outcomes were reported meant that not all studies could be incorporated into meta-analyses (e.g. for exacerbations and some hospitalisation data). Small numbers of studies in meta-analyses (up to seven maximum) precluded the construction of funnel plots, and the small number, together with a lack of reporting of some quality criteria, meant that sensitivity analyses around study quality were also not feasible.

Search strategy

Searches for economic studies were run on MEDLINE, EMBASE via Ovid and NHS Economic Evaluation Database using, where appropriate, relevant terms for economic studies along with terms for clinical populations. Examples of these strategies can be found in Appendix 1. These searches were supplemented with any further economic evaluations and cost studies identified during screening of the search yield in the clinical effectiveness review.

Study selection

All records were screened by two reviewers independently, and copies of potentially relevant articles were obtained for scrutiny against the full selection criteria, with any disagreements resolved by discussion. The inclusion criteria were the same as those for the clinical effectiveness review except that the study designs were full economic evaluations, partial economic evaluations and cost-effectiveness analyses alongside trials and economic modelling studies.

Data extraction and quality assessment strategy

Data on the following, where available, were extracted from included studies by one reviewer and checked by another:

• study characteristics, such as study question, form of economic analysis, population, interventions, comparators, perspective, time horizon and form of modelling used

• clinical effectiveness and cost parameters, such as effectiveness data, health state valuations (utilities), resource-use data, unit-cost data, price year, discounting and key assumptions

• results and sensitivity analyses.

Studies were to be quality-assessed using the Drummond checklist132 for economic evaluations and the checklist by Philips133 for model-based analyses.

Summary of included studies

Both included studies conducted a cost analysis alongside a clinical study of domiciliary NIV in COPD patients. 40,134 Each study met at least eight of the 10 Drummond and Jefferson quality-assessment criteria (see Appendix 11). 132 Neither study was sponsored by industry, although some potential conflicts of interested were declared by Tuggey et al. (2003). 40 The characteristics and main results of each study are summarised in Table 19.

Tuggey et al. (2003)

Tuggey et al. ’s (2003)40 cost analysis was based on a before-and-after study of 13 patients offered domiciliary NIV in one UK hospital between 1995 and 2000. Patients were started on NIV if they had a history of recurrent COPD-related hospital admissions (mean of five per year) and had demonstrated prior tolerance of NIV during treatment of an acute exacerbation. As the clinical study enrolled a highly select group of COPD patients, the results are not necessarily transferable to a wider COPD population.

Outcomes were collected from case notes and as a result are potentially biased because of the risk of reporting errors. The study found a statistically significant reduction in admissions, which was the primary measure of effect considered in the subsequent cost analysis. Costs were collected retrospectively and the cost analysis was conducted from a hospital perspective, and thus did not include costs incurred by patients, carers, primary care or the wider society. Costs included acute and conventional hospital admissions, intensive care treatment, outpatient appointments and the cost of domiciliary NIV. The cost of trialled patients on NIV in hospital was not considered as an additional cost, as this was considered part of usual care.

A number of assumptions regarding the costs of acute and domiciliary NIV were drawn from Plant et al. (2003),135 a published economic evaluation of NIV for managing acute exacerbation in COPD patients. In this analysis the provision of domiciliary NIV was estimated to cost £1060 per patient per year in 2003 NHS prices. This included £570 for ventilator equipment discounted over 5 years, £224 for mask and tubing with an assumed lifespan of 8 months, £179 for a warm air humidifier with an assumed lifespan of 1 year, £28 for annual servicing and £60 for access to a respiratory nurse specialist.

The results of this analysis found domiciliary NIV in COPD patients with recurrent exacerbation, implemented at discharge, to be cost-saving, resulting in a net saving of £8254 (95% CI £4013 to £12,495) per patient per year. A one-way sensitivity analysis was conducted by varying key assumptions from baseline between the 5th and 95th centiles, and in all scenarios domiciliary NIV was cost-saving compared with usual care alone.

In summary, this study suggests domiciliary NIV may be cost-effective in NIV patients with a history of readmission to hospital. Caution should be taken in interpreting these results, as they are based on a very small sample size and highly select group of patients.

Clini et al. (2009)

Clini et al. (2009)134 presented the results of an Italian cost analysis, conducted alongside a RCT of domiciliary NIV, where 77 patients were followed up for 2 years. Patients with stable COPD, with a prior prescription of LTOT, were enrolled on the trial and randomised to NIV or usual care alone.

This cost analysis applied the mean reduction in hospital admissions and length of stay derived from the RCT. It should, however, be noted that, while the outcomes found in the RCT were in favour of NIV, the results applied were not statistically significant. The resources considered in the analysis included hospital admissions, drug therapy, LTOT and domiciliary NIV equipment and training. Neither the other primary-care resources and costs incurred by patients and their families nor the additional hospital and set-up costs to familiarise patients with NIV were included.

The study estimated the cost of domiciliary NIV to be €160 per month or €1920 per year based on 2008 prices. This estimate was based on a contract with a regional health-care provider and included the provision of equipment, a Respironics BiPAP^® ST-30 (Murrysville, PA, USA) device and warm air humidifier, as well as tubing and masks replaced every 6 months. The results suggest that the addition of domiciliary NIV to LTOT results in a similar cost per patient per day of €23.73 ± €16.18, compared with €21.42 ± €20.38 in the group with LTOT alone, from which the authors concluded that domiciliary NIV was likely to be cost-neutral, as the additional costs were offset by savings from reduced hospital admissions and shorter hospital stays.

In summary, this study suggests that domiciliary NIV may be cost-neutral in stable COPD patients. However, as this study was conducted in Italy, the costs and outcomes are not directly transferable to UK patients. The results should also be considered cautiously because of uncertainty around the effect.

Commentary on excluded studies

Two studies excluded from this review warrant comment.

Criner et al. (1995)136 conducted a retrospective cost analysis of providing NIV to patients admitted to a ventilator rehabilitation unit for moderately severe respiratory failure in the USA. There was some ambiguity as to whether or not this met the inclusion criterion, as it was unclear if patients were discharged with NIV for home use. On balance, the decision was made to exclude the study, as it appeared to primarily assess the effect of NIV delivered in an acute setting. This study demonstrated the interdependence between acute and domiciliary NIV services when considering costs and outcomes.

Chandra et al. (2012)137 published a Markov model developed as part of a series of papers to evaluate the effectiveness and cost-effectiveness of interventions in COPD populations sponsored by the Medical Advisory Secretariat, Canada. NIV in stable COPD patients was included in this series and the authors documented their intent to conduct a clinical review and economic evaluation. 137 However, after conducting the clinical review, they concluded that there was no evidence that domiciliary NIV was effective in COPD patients and thus did not conduct an economic analysis.

Model description

A Markov decision model was built in TreeAgePro (TreeAge Software, Inc., 2013) to compare the cost-effectiveness of domiciliary NIV with usual care from a UK perspective.

The model was developed subsequent to a review of published Markov decision models138–147 in stable and post-admission COPD populations, and considers stable and post-admission populations separately. The model incorporated both the short-term increased risk of readmission and subsequent mortality after a hospital admission, and the long-term natural history of the disease, taking into account exacerbations, increasing COPD severity and mortality (Figure 16). Health states are linked to GOLD severity (using the pre-2011 definition of GOLD because of a lack of available data for the current classification). It incorporates evidence on the increased risk of mortality and readmission in those recently discharged after admission for COPD exacerbation. Tunnel states were added to incorporate the increased risk of mortality and readmission in those recently discharged after admission for COPD exacerbation. 129

FIGURE 16.

Model health states.

The model considers stable or post-discharge health states (at months 1, 2 and 3); that is, patients can be a stable condition or in month 1, 2 or 3 of recovery after a recent exacerbation. Patients can move between stable and post-discharge health states, and there is a risk of death while in any health state. The two populations, stable and post-hospital, differ in where they enter the model and in the probability of moving between the different health states.

The model had a time cycle of 1 month and a lifetime time horizon (30 years) was used. All costs and outcomes were considered from a UK NHS perspective for a price year of 2012.

The model had three transition health states and one stable health state, with each state subdivided into two disease severity levels (see Figure 16). The three transition health states captured patients’ elevated risk of readmission or death in each of the 3 months subsequent to a hospital admission and were referred to as the ‘post-discharge’ health states. A post-discharge period of 3 months was chosen as there are consistent data to indicate elevated risk of readmission and death during this period (2011 audit129). The ‘stable’ health state represented patients’ ongoing risks of events (e.g. exacerbation, hospital admission, death) beyond this period. The subdivision of these states represented GOLD stages 3 and 4. At any point patients could transition to a worse disease severity or die. As COPD is a progressive disease, patients could not transition to an improved health state (e.g. from GOLD stage 4 to 3).

The two COPD severity levels considered were defined according to the traditional GOLD classification stages 1–4. GOLD stage 3 (severe COPD) was defined as having a predicted FEV₁ ≥ 30% but < 50% and GOLD stage 4 (very severe COPD) a predicted FEV₁ ≤ 30%. Health states for GOLD stage 1 (mild COPD) and GOLD stage 2 (moderate) were excluded, as it was assumed that the majority of patients with end-stage COPD were at GOLD stage 3 or 4. After any cycle, patients in GOLD stage 3 could move to a GOLD stage 4 health state as their disease progressed, but GOLD stage 4 patients could not move back to GOLD stage 3.

The stable health states for the stable population reflected outcomes (e.g. exacerbation, readmission, death) reported in large cohorts of COPD patients who were stable for at least 12 weeks, and the post-discharge health states reflected short-term outcomes immediately after admission for severe exacerbation.

The stable health states for the post-hospital population were adapted slightly and populated using different sources to reflect outcomes reported in cohorts of COPD patients followed up after admission to hospital. The post-discharge health states for the post-hospital population were identical to those applied in the stable population (further details on parameterisation of the model are given in Estimation of model parameters).

Transitions within the stable health states

In the stable health state, patients could die from a non-COPD-related cause, live exacerbation-free or experience an exacerbation. The exacerbation could be moderate (managed at home) or severe (requiring admission) or be fatal. Those who survived a severe exacerbation were discharged and moved to a first month post-discharge health state. Those who lived without experiencing an exacerbation or experienced a moderate exacerbation re-entered a stable health state. An example of the pathway within a stable health state is illustrated with reference to GOLD stage 3 for the stable population in Figure 17.

FIGURE 17.

Example of pathway within stable health state with reference to GOLD stage 3 for the stable population.

This pathway was almost identical for GOLD stage 4 in the stable population, but patients were not able to move to a GOLD stage 3 health state.

The pathways for the GOLD stage 3 and 4 stable health for the post-hospital population were similar but populated using different sources. As the risk of a moderate exacerbation (managed at home) was not reported for this population, the risk of exacerbation without admission was not considered for the post-hospital population.

Transitions within the post-discharge health states

Patients who entered the post-discharge health state could die at home, continue their recovery or be readmitted, where they could die during admission. If they survived the hospital admission, they re-entered one of the first month post-admission health states. If recovery continued without being readmitted, they moved to the second and then third post-admission health states where they faced similar pathways. The additional costs and utility losses associated with a non-severe exacerbation during the recovery period were considered negligible, as patients were already assumed to have incurred higher costs and utility loss.

The pathways within the post-discharge health states were almost identical for months 1–3, but in month 3 patients could transition to a stable health state and differed only in disease severity, allowing GOLD stage 3 patients to transition to a parallel GOLD stage 4 state. As noted above, the post-discharge health states were identical for the stable and post-hospital populations. An example of the pathway within a post-discharge health state is illustrated with reference to the first month post discharge for GOLD stage 3 in the stable population in Figure 18.

FIGURE 18.

Example of pathway within a post-discharge health state with reference to GOLD stage 3 for the stable population.

Estimation of model parameters

This section outlines in detail the assumptions applied and sources used to populate the base-case parameters for usual care and domiciliary NIV in both populations. For those readers wishing to just read a summary of the assumptions and data sources applied in the model, these are provided in Chapter 7, Assessment of cost-effectiveness.

Resource use and costs

The resource use considered within the model was broadly concerned with primary and secondary health-care professional contacts and pharmacotherapy for usual care and the additional costs associated with domiciliary NIV. Health-care contacts for each GOLD severity state were estimated with reference to NICE guidelines4 and expert opinion. Use of pharmacotherapy was estimated from data provided by the BLISS cohort. 149 Unit costs were primarily obtained from NHS reference costs156 and Unit Costs of Health and Social Care. 157 Where appropriate, unit costs were inflated to 2012 prices using NHS health index inflation rates. Annual costs were divided by 12 to derive a monthly cost. Moderate and severe exacerbations were treated as additional one-off costs and assumed to be the same irrespective of the underlying GOLD stage.

Assessment of cost-effectiveness

The incremental analysis was designed to generate the cost per additional QALY gained for the addition of domiciliary NIV to usual care in two populations (stable and post-hospital), when compared with usual care alone, in a cohort of COPD patients.

Where available, data were entered into the model as distributions in order to fully incorporate the uncertainty around parameter values so that a probabilistic sensitivity analysis could be undertaken. Where distributions were not available for cost estimates, a normal distribution was applied, assuming a 10% variation either side of the point estimate. Beta distributions were applied to the proportion on different treatments and the proportion accessing services in primary and secondary care. Beta distributions were also applied to annual exacerbation rates and the proportion resulting in hospital admissions, as well as the reduced risk of hospitalisation expected in the domiciliary NIV arm. Normal distributions were applied to utilities and utility losses, as well as the utility improvement associated with NIV in the stable population. The probabilistic sensitivity analysis was run with 1000 simulations, and cost-effectiveness planes and cost-effectiveness acceptability curves (CEACs) were produced.

Expected value of perfect information

Expected value of perfect information (EVPI) analysis quantifies the economic value of removing all uncertainty in a decision model. 169 EVPI analyses were conducted based on the methods described by Claxton and Posnett (1996)170 and applied to estimate the value of perfect information per patient and per population in stable and post-hospital COPD populations in the UK.

The sources used to estimate the size of the stable and post-hospital populations are reported in Table 35. The UK COPD population was estimated by applying the prevalence rate of 1.7% reported in Haughney et al. (2013)172 to the mid-year population projections171 for the UK in 2014. The stable COPD population was estimated from this, assuming that 25.5% and 5.2% were in GOLD stages 3 and 4, respectively, as reported in Haughney et al. (2013). 172

The same study also reported the proportion of patients in each of the new GOLD classifications (A–D1) admitted to hospital at least once. This was used to estimate the post-hospital population, assuming that only GOLD stage C and D patients were likely to be in end-stage COPD.

As there is a lot of uncertainty regarding both the prevalence of COPD and the proportion of COPD patients who could be defined as end-stage stable patients and end-stage post-hospital patients, the estimated value of perfect information per population reported should be interpreted cautiously and considered indicative of the value of perfect information in decisions that affect these populations.

Results for the stable population

Base-case analysis

The base-case results for the stable population presented in Table 36 show that, compared with usual care alone, the addition of domiciliary NIV was more costly and resulted in better outcomes. The difference in cost was £12,769, with 0.4534 QALYs gained, resulting in an ICER of £28,162 per QALY gained.

TABLE 36 - Base-case results for the stable population

Strategy	Mean cost (£)	Cost difference (£)	Mean QALYs	QALY difference	ICER (£/QALY)	Probability cost-effective at £20,000/QALY (%)	Probability cost-effective at £30,000/QALY (%)
Usual care	23,969	–	5.3360	–	–	–	–
Domiciliary NIV + usual care	36,737	12,769	5.7894	0.4534	28,162	13	55

Results from the probabilistic sensitivity analysis can be found in the cost-effectiveness plane in Figure 19, which shows the distribution of 1000 resampled cost and effect difference pairs. In all samples, domiciliary NIV was more costly than usual care, but there was uncertainty regarding its effect. The samples to the left of the line of no effect represent where domiciliary NIV results in more admissions and higher mortality, in which case usual care is more favourable. The majority of the samples are to the right of the line of no effect where cost-effectiveness is determined by the decision-maker’s willingness-to-pay threshold for a QALY gained. The CEAC in Figure 20 shows that domiciliary NIV has a 55% probability of being cost-effective at a threshold of £30,000 per QALY gained.

FIGURE 19.

Cost-effectiveness plane for NIV vs. usual care in the base case for a stable population.

FIGURE 20.

Cost-effectiveness acceptability curve for NIV vs. usual care in for the base case for a stable population.

Sensitivity analysis for the stable population

This section presents the results of the sensitivity analysis conducted on the base-case assumptions for the stable COPD population.

Alternative rate ratios for hospital admissions

Owing to considerable uncertainty around the estimate for the primary outcome, the rate ratio for hospital admission, this parameter was expected to be the biggest driver of the cost-effectiveness results. High and low rate ratio estimates were applied, obtained from the RCTs showing the most positive and negative effect on admissions, and the results are reported in Table 37. The most favourable rate ratio increased the difference in cost and effect to £13,593 and 0.7285 QALYs, respectively, and the ICER was £18,660 per QALY gained, increasing the probability of domiciliary NIV being cost-effective at a threshold of £30,000 per QALY gained to 82%.

TABLE 37 - One-way sensitivity analysis in the stable population varying the rate ratio for effect

Rate ratio	Cost difference (£)	QALY difference (QALY)	ICER (£/QALY)	Probability cost-effective at £20,000/QALY (%)	Probability cost-effective at £30,000/QALY (%)
Base
Pooled estimate	12,769	0.4534	28,162	13	55
High estimate
Zhou et al. 2008,81 0.6219 (0.1425)	13,593	0.7285	18,660	57	82
Low estimate
Kaminski et al. 1999,101 1.794 (0.6031)	8735	–1.0507	Usual care dominates	< 1	1

Conversely, applying the rate ratio that marginally favoured usual care decreased the difference in costs to £8735, and domiciliary NIV was less effective than usual care, resulting in QALY difference of –1.0507 QALYs. Under this assumption, usual care dominated NIV and the probability of this being cost-effective reduced to 1% at a threshold of £30,000 per QALY gained.

The ICERs obtained from varying the rate ratios between the low and most favourable estimates are illustrated in Figure 21. The points where the trend line is not on the graph represent ICERs above £60,000 or below –£60,000 per QALY gained and the negative values signify scenarios where usual care dominates domiciliary NIV (usual care costs less and is more effective). For the ICER to be below a threshold of £30,000 per QALY gained, the rate ratio for hospital admissions needs to be 0.76 or lower; that is, a 24% reduction in rate of admissions per patient per year with NIV.

FIGURE 21.

One-way sensitivity analysis of the rate ratio for admissions in a stable population.

Alternative improvements in utility

The assumption that domiciliary NIV has no effect on utility (outside of utility loss avoided through lower risk of admission) was tested by varying an assumption that NIV would lead to a change in utility score of 0.025 (on a QoL scale of 0 to 1) in either direction. Although not directly mappable, this could be considered comparable to a 2.5 point variation on the SGRQ (scale 0–100 points). The effect this utility variation has on the ICER is illustrated in Figure 22. Assuming that domiciliary NIV improved QoL and increased the utility score by 0.025, the ICER decreased to £20,462, close to the threshold of £20,000 per QALY gained. Conversely, assuming that domiciliary NIV decreased patients’ QoL, reducing the utility score by 0.025 increased the ICER to above £40,000/QALY.

FIGURE 22.

One-way sensitivity analysis of QoL for the stable population.

Alternative duration of effect

The duration of the effect of domiciliary NIV in reducing the risk of hospital admission and improving utility was varied from 10 years to 2, 5, 20 and 30 years, with 30 years representing the lifetime of the population. Table 38 shows that assuming that the effect lasted for only 2 years results in an ICER of £93,091 per QALY gained and assuming it lasted for 5 years results in an ICER of £43,510 per QALY gained, with the probability of this being cost-effective at a threshold of £30,000 per QALY gained being less than 0% and 12%, respectively.

TABLE 38 - One-way sensitivity analysis in the stable population varying the duration of effect

Duration of effect	Cost difference (£)	QALY difference	ICER (£/QALY)	Probability cost-effective at £20,000/QALY (%)	Probability cost-effective at £30,000/QALY (%)
2 years	11,865	0.1275	93,091	0	0
5 years	12,328	0.2833	43,510	10	12
Base case (10 years)	12,769	0.4534	28,162	13	55
20 years	12,944	0.5817	22,252	37	73
30 years	12,888	0.5969	21,592	42	75

Conversely, assuming that these effects lasted for 20 years or more reduced the ICER towards £20,000 per QALY gained, and the probability of domiciliary NIV being cost-effective at a threshold of £30,000 per QALY gained was greater than 70%.

The effect of varying the duration of effect between 2 years and 30 years is illustrated in Figure 23. This shows that the ICER decreases at longer durations of effect but, even assuming that the effects last for up to 30 years, the ICER is marginally higher than £20,000 per QALY gained.

FIGURE 23.

One-way sensitivity analysis of the duration of the effects.

Alternative model time horizon

Table 39 presents the results of the model when varying the time horizon of the model. At a short time horizon of 6 months, the ICER was above £475,000 per QALY gained, and using a time frame of 2 years it gave an ICER above £116,000 per QALY gained, both well above the threshold of £30,000 per QALY gained. While these are unlikely to represent realistic time frames, there were very few studies that evaluated domiciliary NIV in this population beyond 2 years; therefore, projecting costs and outcomes beyond this point is associated with uncertainty.

TABLE 39 - One-way analysis in the stable population varying the time horizons

Time horizons	Cost difference (£)	QALY difference	ICER (£/QALY)	Probability cost-effective at £30,000/QALY (%)	Probability cost-effective at £60,000/QALY (%)
6 months	928	0.0019	476,982	0	0
2 years	2346	0.0201	116,857	0	0
5 years	4897	0.0908	53,910	0	5
10 years	8273	0.2526	32,757	7	41
20 years	12,604	0.4466	28,222	12	55
Base case (30 years)	12,769	0.4534	28,162	13	55

There was very little difference between applying a time horizon of 20 years and one of 30 years, as the base case assumed that domiciliary NIV was effective for 10 years, and beyond this the monthly costs would continue to accrue with no further benefit other than continued survival. Thus, at all time horizons the ICER is above £30,000 per QALY gained and the probability of being below £30,000 per QALY gained is less than 55%.

Alternative continuation rates

Varying the assumption that 85% of those started on NIV would benefit from and continue using NIV beyond 3 months between 55% and 95% (representing low and most favourable estimates from selected studies) suggested that the model was not very sensitive to this variable, as illustrated in Figure 24. While a high dropout rate was expected to increase the costs and lower QALYs gained in the short term, these costs were minimal compared with the costs and benefits that accrued in the population that continued with NIV beyond 3 months.

FIGURE 24.

One-way sensitivity analysis of alternative continuation rates beyond 3 months in the stable population.

Alternative costs of the non-invasive ventilation device

In the base case, it was assumed that the device would cost £3600 and a monthly cost was derived assuming that it would be used continuously for 5 years. Varying the cost of the NIV device demonstrated that, if over £4000, the ICER is above the threshold of £30,000 per QALY gained, as illustrated in Figure 25 (keeping the lifespan constant).

FIGURE 25.

One-way sensitivity analysis of alternative costs of device in a stable population.

Keeping the cost of the device constant at £3600, the ICER rises above £30,000 per QALY gained when the device is assumed to be used for 4 years (rather than 5 years), as illustrated in Figure 26.

FIGURE 26.

One-way sensitivity analysis of alternative lifespan of device in a stable population.

Alternative costs of the service aspects of non-invasive ventilation provision

The other costs involved in the provision of a domiciliary NIV service were tested by varying the tariff for a NIV assessment and set-up (base case £354) and the cost of annual maintenance (base case £550) between £200 and £1000. Changing the tariff for NIV assessment and set-up had minimal impact on the cost-effectiveness of NIV, as this was a one-off cost incurred in the first 3 months of use (Figure 27). The model was more sensitive to changes in the annual service cost; however, even when applying a very low cost of £200 the ICER was above £20,000 per QALY gained (Figure 28).

FIGURE 27.

One-way sensitivity analysis of alternative NIV assessment costs in a stable population.

FIGURE 28.

One-way sensitivity analysis of annual maintenance costs in a stable population.

Alternative assumption for cost of admission

The results of varying the cost of an admission from the base-case assumption of £2053 between values of £1500 and £3000 are shown in Figure 29 and demonstrate that the model was not very sensitive to changes in this cost.

FIGURE 29.

One-way sensitivity analysis of cost of admission in a stable population.

Alternative utility assumptions applied to the stable population

The difference in costs and QALYs when using the utility assumptions applied in Borg et al. (2004)138 are reported in Table 40. As these were similar to the values in the base case, it is unsurprising that this made very little difference to the ICER and cost-effectiveness of domiciliary NIV.

TABLE 40 - One-way sensitivity analysis in the stable population using alternative utility assumptions

Strategy	Mean cost (£)	Cost difference (£)	Mean QALYs	QALY difference	ICER (£/QALY)	Probability cost-effective at £20,000/QALY (%)	Probability cost-effective at £30,000/QALY (%)
Usual care	23,969	–	5.3330	–	–	–	–
Domiciliary NIV + usual care	36,737	12,769	5.7888	0.4558	28,013	14	55

Expected value of perfect information for the stable population

The EVPI per patient for the stable population was estimated to be £204 at a threshold of £20,000 per QALY gained and £1812 per patient at a threshold of £30,000 per QALY gained. When this is multiplied by the estimated population size (329,000 stable end-stage COPD patients in the UK), it results in an EVPI per population estimate of £67M and £596M respectively. Figure 30 shows the EVPI per population at all willingness-to-pay thresholds as between £0 and £60,000 per QALYs gained. These estimates reflect the value of research that removes all uncertainty in the decisions to adopt NIV at these willingness-to-pay thresholds. The high value at all thresholds above £20,000 per QALY gained reflects the fact that removing all uncertainty in the decision to offer domiciliary NIV to all end-stage COPD potentially affects a large patient base.

FIGURE 30.

Expected value of per information per population for the stable COPD population.

At a threshold of £20,000 per QALY gained, usual care is the preferred option. The EVPI per population increases above the threshold of £20,000 per QALY gained and reaches a peak at around a little over £28,000/QALY gained, where the decision changes and domiciliary NIV becomes the preferred option. This sharp increase is because, at thresholds close to the base-case ICER, perfect information becomes more important, as it is likely to change the decision to offer domiciliary NIV.

Results for the post-hospital population

This section presents the results of the base-case, sensitivity and subgroup analyses for the post-hospital population.

Base-case analysis

The three base-case results for the post-hospital population presented in Table 45 show that, compared with usual care alone, the addition of domiciliary NIV was more costly and resulted in better outcomes where the rate ratios favoured NIV and resulted in higher costs and worse outcomes with the rate ratio that favoured usual care; therefore, domiciliary NIV was dominated. Where the evidence marginally favoured NIV, the ICER was £10,107 per QALY gained and this reduced to £6281 per QALY gained where the rate ratio strongly favoured NIV. Therefore, both of the rate ratios that favoured NIV produced ICERs that were well under the willingness-to-pay threshold of £20,000 per QALY gained; however, if NIV is not more effective than usual care, it would not be considered cost-effective irrespective of the willingness-to-pay threshold value.

TABLE 45 - Base-case results for the post-hospital population

Strategy	Mean cost (£)	Cost difference (£)	Mean QALYs	QALY difference	ICER (£/QALY)	Probability cost-effective at £20,000/QALY (%)	Probability cost-effective at £30,000/QALY (%)
Usual care	17,048	–	2.1452	–	–	–	–
Domiciliary NIV + usual care
Marginally favoured usual care (Struik et al.75)	21,912	4864	1.8196	–0.3255	Dominated	0	0
Marginally favoured NIV (Cheung et al.90)	22,879	5830	2.722	0.5769	10,107	72	79
Most favourable (Xiang et al.92)	23,533	6485	3.177	1.0325	6281	100	100

The results of the probabilistic sensitivity analysis, showing the distribution of 1000 resampled cost–effect difference pairs for each base case, are shown on three cost-effectiveness planes in Figure 31.

FIGURE 31.

Cost-effectiveness plane of NIV vs. usual care in the three base-case post-hospital populations. (a) Base case marginally favours control; (b) base case marginally favours NIV; and (c) base case most favourable to NIV.

In all three base cases, domiciliary NIV was more costly; however, the results differed widely across the three rate ratios with respect to the effect. For the base case that used evidence that marginally favoured usual care, the majority of the samples are on or to the left of the line of no effect, but there was some uncertainty around this. Similarly, for the base case that used evidence that marginally favoured NIV, in most samples domiciliary NIV was more effective, but there was a lot of uncertainty regarding effectiveness. In the most favourable base case, NIV was more costly and more effective in all samples.

The CEAC in Figure 32 show that, at a threshold of £20,000 per QALY gained, domiciliary NIV has a 0%, 72% and 100% probability of being cost-effective in the base case that marginally favoured usual care, the base case that marginally favoured NIV and the base case that was most favourable to NIV, respectively. This suggests that the considerable uncertainty around the effect of NIV on admissions follows through into the cost-effectiveness analysis findings.

FIGURE 32.

Cost-effectiveness acceptability curve for NIV vs. usual care in the three base-case post-hospital populations.

Sensitivity analysis

This section presents the results of the sensitivity analysis conducted on the base-case assumptions applied to the post-hospital COPD population.

Alternative rate ratios for hospital admissions

Varying the rate ratio for admission in this population between 0.05 and 0.95, as illustrated in Figure 33, shows that the ICER for NIV is very sensitive to this parameter. At rate ratios of 0.85 and above (i.e. a 15% reduction in the rate of admissions per patient per year with NIV), the ICER is above £30,000 per QALY gained, and at rate ratios of 0.75 and below, the ICERs are below £20,000 per QALY gained.

FIGURE 33.

One-way sensitivity analysis varying the rate ratio for admissions in a post-hospital population.

Alternative improvements in utility

While there was no evidence to suggest that domiciliary NIV had an effect on utility, the assumption that domiciliary NIV may have a positive or negative effect on QoL was tested by applying a change of 0.025 in the utility score (on a scale of 0 to 1) in either direction to the three base cases. The effects on the ICERs are illustrated in Figure 34. Assuming that NIV had a small positive or negative effect on the utility score (by 0.025), this had little impact on cost-effectiveness. Usual care dominated NIV in all cases in which the base case marginally favoured usual care. Conversely, where the base cases favoured domiciliary NIV, in all cases the ICER remained below £20,000 per QALY gained.

FIGURE 34.

Alternative assumption for effect on utility in a post-hospital population.

Alternative durations for effect

Figure 35 presents the results, varying the duration of the effect relating to the three base cases from 10 years to between 1 and 30 years. In both of the base cases in which the effectiveness estimate favoured NIV, the ICERs were below £20,000 if the effect of NIV lasted for 2 years or more. Conversely, in the base case in which the evidence marginally favoured usual care, changing the duration of this effect, unsurprisingly, did not change the likelihood of NIV being cost-effective and, in all cases, usual care dominated NIV.

FIGURE 35.

One-way analysis changing the duration of effect in a post-hospital population.

This suggests that it is important to verify the effects of NIV beyond 2 years, as this assumption is an important determinant of the cost-effectiveness of this intervention in this population, but it is less important if NIV is effective beyond 5 years in this population.

Alternative model time horizons

Table 46 presents the results when varying the time horizon of the model. Usual care dominated NIV for all possible time horizons. With a time horizon of 6 months, the ICERs for marginally favouring NIV and strongly favouring NIV were £59,775 per QALY gained and £18,805 per QALY gained, respectively. After 2 years, the ICERs decreased to £14,806 per QALY gained and £2023 per QALY gained, both with an increasing probability of being cost-effective. At all time horizons beyond 5 years, the ICER was £10,107 per QALY gained or less for both the cases that favoured NIV and there was only a small change to the probability of domiciliary NIV being cost-effective, varying between 75% and 80% at a threshold of £30,000 per QALY gained in the case that marginally favoured NIV and staying at 100% in the most favourable case.

TABLE 46 - One-way sensitivity analysis varying the time horizon in a post-hospital population

Direction of effect	Time horizon	Cost difference (£)	QALY difference	ICER (£/QALY)	Probability cost-effective at £20,000/QALY (%)	Probability cost-effective at £30,000/QALY (%)
Marginally favoured usual care	6 months	1289	–0.0057	Dominated	0	0
	2 years	3164	–0.0465	Dominated	0	0
	5 years	4719	–0.1476	Dominated	0	0
	10 years	5085	–0.2655	Dominated	0	0
	20 years	4881	–0.3209	Dominated	0	0
	30 years (base case)	4865	–0.3255	Dominated	0	0
Marginally favoured NIV	6 months	504	0.0084	59,775	17	31
	2 years	995	0.0672	14,806	60	68
	5 years	2301	0.2277	10,106	67	75
	10 years	4119	0.4509	9135	69	77
	20 years	5697	0.5673	10,043	72	78
	30 years (base case)	5830	0.5769	10,107	71	77
Most favourable	6 months	249	0.0132	18,805	53	80
	2 years	221	0.1092	2023	98	100
	5 years	1285	0.3828	3357	100	100
	10 years	3581	0.7918	4523	100	100
	20 years	6259	1.0141	6171	100	100
	30 years (base case)	6485	1.0325	6281	100	100

Similar to varying the duration of effect, the model was less sensitive to changes in the time-horizon beyond 2 years compared with the stable population because of higher baseline all-cause and COPD-related mortality risks.

Alternative continuation rates

The effect on the ICER of varying the proportion of those who benefit from domiciliary NIV and continue using it from 55% to 95% in the three base cases is shown in Figure 36. As with the stable population, the model was not very sensitive to this assumption as the short-term costs of trialled domiciliary NIV on those who discontinue without benefiting was minimal compared with the costs and benefits that accrue over the lifetime of the population that continues with NIV.

FIGURE 36.

One-way sensitivity analysis of alternative continuation rates beyond 3 months in a post-hospital population.

Alternative costs of the non-invasive ventilation device

In the base cases it was assumed that the NIV device would cost £3600 and the monthly cost was derived on the assumption that it would be used continuously for 5 years. The results of varying the cost of the NIV device are illustrated in Figure 37 and show that at all costs of the device between £2000 and £6000 the ICER was below the lower threshold of £30,000 per QALY gained for the base cases in which the effectiveness estimate favoured NIV, and was dominated at all costs for the base case that favoured usual care.

FIGURE 37.

One-way sensitivity analysis of alternative costs of device in a post-hospital population.

Similarly, the assumption for the lifespan of the device was varied between 4 and 8 years and, as illustrated in Figure 38, in all cases the ICER was below £30,000 per QALY gained for the base cases in which the effectiveness estimate favoured NIV, and was dominated at all lifespans for the case that favoured usual care.

FIGURE 38.

One-way sensitivity analysis of alternative lifespan of device in a post-hospital population.

Alternative costs of the service aspects of non-invasive ventilation provision

The tariff for an NIV assessment and set-up varied between £200 and £1000. The results in Figure 39 show that the model was not sensitive to variation in this cost in any of the cases, as it was a one-off cost and had a minimal impact on the cost difference over the cohort’s lifetime. The model was more sensitive to changes to the annual service cost (Figure 40). However, as with the cost of the NIV device, the ICERs were below £15,000 per QALY gained for the two base cases in which the effectiveness estimate favoured NIV and were dominated in all cases where usual care was favoured.

FIGURE 39.

One-way sensitivity analysis of alternative NIV assessment costs in a post-hospital population.

FIGURE 40.

One-way sensitivity analysis of annual maintenance costs in a post-hospital population.

Alternative utility assumptions applied in the post-hospital population

The difference in costs and QALYs when using the utility assumptions applied in Borg et al. (2004)138 are reported in Table 47. Similarly to when these were applied to the stable population, this made very little difference to the cost-effectiveness of domiciliary NIV in any of the three post-hospital population base cases.

TABLE 47 - One-way sensitivity analysis in the post-hospital population using alternative utility assumptions

Strategy	Mean cost (£)	Cost difference (£)	Mean QALYs	QALY difference	ICER (£/QALY)	Probability cost-effective at £20,000/QALY (%)	Probability cost-effective at £30,000/QALY (%)
Usual care	17,048	–	2.2715	–	–	–	–
Domiciliary NIV + usual care
Marginally favoured usual care	21,912	4864	2.1614	–0.3294	Dominated	0	0
Marginally favoured NIV	22,879	5830	2.7447	0.5833	9996	72	79
Most favourable	23,533	6485	3.2014	1.04	6236	0	0

Alternative baseline risks applied in the post-hospital population

Replacing the parameters for hospitalisation and mortality obtained from Garcia-Aymerich et al. (2003)152 with those obtained from Bucknall et al. (2011)153 increased the ICER to £23,211 per QALY gained in the case in which the effectiveness estimate marginally favoured NIV and to £13,101 per QALY gained with the most favourable effectiveness estimate (Table 48). It reduced the probability of NIV being cost-effective at a threshold of £30,000 per QALY gained to 61% in the case that the effectiveness estimate marginally favoured NIV, but had no effect on the probabilities of the other cases. This suggests that domiciliary NIV is less cost-effective in patients with lower baseline risk of hospitalisation and/or mortality.

TABLE 48 - One-way sensitivity analysis of alternative baseline risks in a post-hospital population

Severity	Cost difference (£)	QALY difference	ICER (£/QALY)	Probability cost-effective at £20,000/QALY (%)	Probability cost-effective at £30,000/QALY (%)
Applying baseline risks from Bucknall et al.153
Marginally favoured usual care	5613	–0.1346	Dominated	0	0
Marginally favoured NIV	4034	0.1738	23,211	34	61
Most favourable	3562	0.2719	13101	99	100
Doubling all mortality risks
Marginally favoured usual care	2567	–0.1637	Dominated	0	0
Marginally favoured NIV	3019	0.3147	9593	75	80
Most favourable	3379	0.587	5756	100	100

Adjusting all mortality risks (risk of death during exacerbation, risk of death post-exacerbation and risk of all-cause mortality) by a factor of two decreased both the costs and outcomes compared with the base case, reflecting shorter survival in both arms. This reduced the ICER to £9593 per QALY gained in the case where the effectiveness estimate marginally favoured NIV and £5756 per QALY gained with the most favourable effect estimate and had a minimal impact on the likelihood of domiciliary NIV being cost-effective at £20,000 or £30,000 per QALY gained. For all analyses, usual care remained dominant over NIV with the case that used an effect estimate that favoured usual care.

Expected value of perfect information analysis for the post-hospital population

The EVPI per patient for the post-hospital population was estimated to be £1541 and £1694 at a threshold of £20,000 and 30,000 per QALY gained, respectively, for the case in which effectiveness evidence marginally favoured NIV. This reflects the high value of removing the uncertainty in this case, in which there was considerable uncertainty around the likelihood of NIV being cost-effective at these thresholds. Conversely, the value of EVPI for the case in which the effectiveness estimate that marginally favoured usual care and in which the intervention was most favourable for NIV was zero in both cases, as the probability of NIV being cost-effective was 0% and 100%, respectively.

Multiplying the EVPI per patient to the estimated population size (77,000 post-hospital patients in the UK) generated an estimate for the EVPI per population of £119M and £130M, respectively, for the case that marginally favoured NIV. The EVPIs per population at thresholds between £0 and £60,000 per QALY gained are illustrated in Figure 41. This EVPI curve reaches a peak at thresholds close to the base-case ICER, at which point the decision changes and domiciliary NIV becomes the preferred option. These estimates reflect the value of conducting research that would remove all uncertainty from this decision and are high because of the uncertainty around the effect and the potentially large population that would be affected by the decision to adopt this technology (albeit a smaller population than all stable end-stage COPD patients).

FIGURE 41.

Expected value of perfect information analysis for the three post-hospital populations.

In the case where effectiveness evidence marginally favoured usual care, the EVPI decreases between £0 and £6000 per QALY gained, then remains at zero until the £30,000 threshold. The EVPI increases above this point, reflecting the small probability that NIV results in higher costs and better outcomes.

At most willingness-to-pay thresholds the EVPI for the case that most favours NIV is zero, reflecting the higher probability of NIV being cost-effect at all thresholds considered. There is a small peak around the point estimate, suggesting that, if the willingness-to-pay per QALY gained threshold were close to £6000, there would be some value to further research of knowing with greater certainty that NIV was cost-effective.

Stable population

Applying the assumption that domiciliary NIV results in a reduction in hospital admissions, the base-case results for the stable population suggest that domiciliary NIV may be a cost-effective intervention at a threshold of £30,000 per QALY gained. However, probabilistic sensitivity analysis estimates that the probability of being cost-effective at this threshold is only 55%, demonstrating the uncertainty around the impact of domiciliary NIV on admission and utility in this population. If a £20,000/QALY gained threshold is applied, the intervention is not cost-effective. The impact of a reduced risk of admission led to lower mortality and morbidity; however, as these effects were expected to last for 10 years and the costs were assumed to accumulate over the cohort’s lifetime, it is uncertain whether or not benefits outweigh the higher lifetime costs of domiciliary NIV provision. The one-way sensitivity analyses undertaken were informative in highlighting the key drivers of the model results. As expected, cost-effectiveness was affected by the estimate of effect on admissions and utility improvement, the duration of effect and elements of the cost of domiciliary NIV provision. Applying the high effectiveness estimate reduced the ICER to £18,660 per QALY gained, and the probability of NIV being cost-effective at £30,000 per QALY gained increased to 82%. However, when applying the effectiveness estimate (which marginally favoured usual care), the intervention was dominated by usual care and this probability was reduced to less than 1%, highlighting the uncertainty around the effect. Varying the rate ratio for hospital admissions found that for NIV to be cost-effective at this threshold, the rate ratio would need to be 0.76 or lower (i.e. a 24% reduction in the rate of admissions per patient per year with NIV). Similarly, assuming domiciliary NIV improved QoL and increased the utility score by 0.025, the ICER decreased to £26,462, closer to the threshold of £20,000 per QALY gained.

It is unsurprising that the intervention becomes more cost-effective when extending the duration of effect beyond 10 years, and is less effective at shorter durations. While it is reasonable to conclude that the effects are likely to last beyond those reported in trials (1 or 2 years) the model becomes more speculative when assuming that the effects extend beyond 10 years. Taking into account all the uncertainty around the effect estimate for hospital admission and effect on utility, it is plausible that the intervention may be cost-effective at £30,000 per QALY gained if the effects lasted for a patient’s lifetime. As the base-case ICER was between the willingness-to-pay thresholds of £20,000 and £30,000 per QALY gained, the model was sensitive to costs that accumulated over the long term, such as the monthly equipment costs and annual maintenance costs. If these costs were lower than those applied in the base-case analysis, the ICER comes closer to £20,000 per QALY gained; however, the overall uncertainty around the results would remain.

The population EVPI was £596M, which reflects the value of removing all uncertainty regarding the decision to adopt domiciliary NIV at a willingness-to-pay threshold of £30,000 per QALY gained. This value is high because of the large population potentially affected by this decision, and should be considered indicative because of uncertainties regarding the prevalence of COPD and the proportion considered end stage and stable.

Post-hospital population

Three base cases were explored in order to reflect the considerable uncertainty of the effect of NIV on hospital admissions. For the case in which the effectiveness estimate marginally favoured usual care, NIV was more costly and resulted in worse outcomes; therefore, usual care was dominant. While there was some uncertainty around the effect, there was a zero likelihood of NIV being cost-effective, and this conclusion was not sensitive to any of the other assumptions tested in sensitivity analysis.

Conversely, the opposite conclusion could be drawn from the case in which the effectiveness estimate was most favourable to NIV. Here NIV was more costly but more effective and, as the ICER was below £10,000 per QALY gained, there was a 100% probability of NIV being cost-effective at a threshold of £20,000 per QALY gained. The case in which the effectiveness estimate marginally favoured NIV also produced an ICER close to £10,000 per QALY, but there was a lot more uncertainty around the probability of NIV being cost-effective, driven primarily by the uncertainty around the effect.

The disparity in the findings from the three base cases highlights the need for more studies to evaluate the effect in this population. Although two out of the three studies informing the base cases favoured NIV and thus found NIV to be cost-effective, the sample size was very small in both cases (n = 40 and 47) and there was also a large difference in size of effect. The study that marginally favoured usual care was larger (n = 201), but the direction of effect was not consistent with that of the other two RCTs.

The results from the EVPIs conducted for each case, not surprisingly, also gave very mixed values for perfect information to inform the decision to offer domiciliary NIV to this population. The value of perfect information in the case in which the effectiveness estimate marginally favoured usual care (obtained from a large study) and the case in which the effectiveness estimate favoured NIV (obtained from a small study) was much smaller than the value of perfect information in the case in which the effectiveness estimate marginally favoured NIV. This was associated with a lot of uncertainty. As a pooled rate ratio for hospital admissions is likely to be closer to the middle estimate (i.e. marginally favouring NIV) than the other two cases, this suggest that there is a strong argument in favour of conducting further research to remove this uncertainty.

Sensitivity analyses

The sensitivity analysis conducted for both populations found that the relative cost-effectiveness of domiciliary NIV was not particularly sensitive to the initial costs of set-up or to the proportion that discontinued using domiciliary NIV in the first 3 months (assuming that only those who benefited continued using it and incurred costs beyond this point). Conversely, the model was sensitive to changes in the size of the effects (rate ratio for admission and change in utility score) and duration of these effects in both populations.

In the base cases for both populations it was assumed that there was no improvement in QoL associated with domiciliary NIV outside of the short-term decrements to QoL avoided from a reduced risk of readmission. This was based on limited QoL data being reported in the studies identified. As this was a conservative assumption, it may have underestimated the relative cost-effectiveness of domiciliary NIV and points to the importance of gathering more robust evidence QoL.

The one-way sensitivity analysis varying the baseline mortality risks applied to the usual-care arm in both populations found that the model was sensitive to this variable and that domiciliary NIV was more cost-effective in patients at greater risk of admission and death. This is logical, demonstrating that if domiciliary NIV is effective in reducing admissions and associated mortality, it will be more effective and cost-effective in patients most at risk of these events. This also highlights the importance of collecting more robust data on patient characteristics that determine the risk of admission and mortality. It also suggests that prioritising those most at risk of these events within NIV services could be considered.

Strengths and limitations

A key strength of this analysis is that it is the first economic model to consider the cost-effectiveness of domiciliary NIV in both stable and post-hospital COPD patients and illustrates the key variables that impact on the results. A further strength of this study was the model structure applied, which is a modified version of previously published decision model where additional post-admission health states were added to incorporate evidence on the higher risks in COPD patients immediately after discharge. This was particularly useful when modelling two different COPD populations with different baseline risks of admission and mortality.

Although there was a great deal of uncertainty around effectiveness data and assumptions applied to this model, distributions were applied to reflect this uncertainty. A probabilistic sensitivity analysis was undertaken, and this was supplemented with an extensive one-way sensitivity analysis of key parameters, thus demonstrating which parameters were mostly likely to influence decisions to implement domiciliary NIV in each patient population. Furthermore, value of information analysis was conducted to quantify the value of conducting further research to eliminate this uncertainty.

However, caution should be applied when interpreting the results of the analyses for both populations modelled. This is a speculative decision model and should only be considered as indicative of the potential cost-effectiveness of domiciliary NIV if it is effective at reducing hospital admissions. For the stable population, there was some evidence of a reduced risk of admissions in some studies reporting this outcome, but there was no evidence of a reduced risk of mortality (up to 2 years). There was also a lot of uncertainty for the post-hospital population, as studies reported very different effect estimates for hospital admissions. A further limitation is the lack of long-term admissions data (up to 2 years). However, both models have assumed an effect on admissions over the first 10 years.

In addition to the uncertainty around the effect of domiciliary NIV, there was also uncertainty around parameters and assumptions for usual care in both populations. The model was populated using some trial evidence but, as the studies identified in the clinical effectiveness systematic review reported limited data, parameters were additionally obtained from large cohort studies. Furthermore, while the model was able to reflect mortality and readmission risks in the first 3 months after discharge, it was assumed that, after those 3 months, those who were not readmitted would move to a stable health state. The baseline admission and mortality rates were assumed to be different for each population. In the post-hospital group, these risks were obtained from a cohort study that followed up patients admitted for an exacerbation. 152 In the stable population, data were extracted from the TORCH148 and Eclipse2 studies, representing average exacerbation and hospitalisations rates in stable COPD cohorts over a 3-year period, and these data were applied over a 30-year time horizon. In reality, these rates may vary over time in both populations.

While a relatively simple EVPI analysis was conducted, which reflected the value of perfect information across all parameters, as opposed to a more complex analysis of perfect information associated with specific parameters, the high values (>£100M) for both populations at £30,000 per QALY gained are indicative of the value of conducting further research to remove all uncertainty from adoption decisions. This suggests that it is worth spending money on research up a maximum of these high values in order to further inform all the model parameters, and eliminate uncertainty. Conducting further expected value of perfect parameter information analysis would help quantify the value of removing uncertainty around particular parameters, but it is expected that continued evaluation of the effect of domiciliary NIV on admissions, mortality and QoL should be a research priority for both populations. Furthermore, the values for EVPI per population analyses reported assumed a COPD prevalence rate of 1.7%171 and that 30.8% and 8.5% of COPD patients would meet the criteria of being end stage and stable, and end stage and post hospital, respectively. 172 As there is a lot of uncertainty around both the true prevalence of COPD and how end-stage patients should be defined, the values for the EVPI per population reported should be interpreted cautiously and considered indicative of the value of perfect information to inform decisions that affect these populations.

The model highlights a number of areas where further research is required. Crucially, further evidence is needed on the effectiveness of domiciliary NIV in reducing admissions and improving QoL. There is a lot of uncertainty around the effectiveness and cost-effectiveness of domiciliary NIV in a stable population, and the intervention may be cost-effective at £30,000 per QALY gained when applying current evidence, but with only a 55% probability of being cost-effective. If NIV is effective in reducing the risk of admissions in a post-hospital population, then it is likely to be below the willingness-to-pay threshold of £20,000. This must be tempered with the fact that the largest study undertaken so far in this population found no benefit of NIV in terms of hospital admissions. More evidence is required on what patient characteristics predict exacerbations, hospitalisation and mortality in COPD populations. It is clear that these events differ by severity of COPD and by recent experience of a hospital admission, although more data are needed on how these two variables interact over time. Finally, while utility values in COPD populations are relatively consistent by GOLD severity, there is a dearth of data available on the impact of exacerbations on QoL.

Summary

• Currently, there are no published economic models on the cost-effectiveness of domiciliary NIV versus usual care in end-stage COPD either starting on NIV when stable in the community or at discharge when recovering from an exacerbation.

• This is the first economic model to attempt to estimate the cost-effectiveness of domiciliary NIV in these patient groups.

• This speculative model indicates that domiciliary NIV may be cost-effective in the stable population at a threshold of £30,000 per QALY gained, but there is considerable uncertainty around this result. Changing parameter estimates around the duration of effect, cost of NIV and target group (e.g. stable patients in GOLD stage 4 or those with a higher risk of repeat exacerbation and mortality) reduces the ICER further to values closer to £20,000 per QALY gained. However, there is a lot of uncertainty around these results.

• There is a lot of uncertainty around the cost-effectiveness of NIV in the post-hospital population. Two cases based on effectiveness estimates from single RCTs with small sample sizes indicated that domiciliary NIV might be cost-effective in a post-hospital population at a willingness-to-pay threshold of £30,000 per QALY gained. This is consistent with the findings above, that if NIV is more effective in reducing the risk of admission it is more likely to be cost-effective in patients with a higher risk of admission and mortality. However, the opposite conclusion was drawn from a third base case, which used data from a larger RCT that favoured usual care with regard to the effect on hospitalisations. More evidence is thus required to investigate the reason for this disparity in outcome between studies in this population.

• The model has a number of limitations, the most important relating to the large amount of uncertainty around the effectiveness estimate driving the model results.

• The analysis highlights the importance of conducting further research on the effect and duration of effect of domiciliary NIV, and on whether or not a population of COPD patients not likely to benefit can be identified, to allow a more robust analysis of cost-effectiveness.

• It also highlights the importance of estimates of baseline risk of admission and COPD-related mortality when assessing the cost-effectiveness of NIV.