Representation of people with comorbidity and multimorbidity in clinical trials of novel drug therapies: an individual-level participant data analysis

Drug treatments that have been recommended in evidence-based clinical guidelines are less likely to be prescribed to people with multimorbidity (defined as people with two or more conditions) [1‐5]. One reason for this difference in prescribing is that the populations included in clinical trials, which underpin evidence-based guidelines, are believed to be unrepresentative of people with multimorbidity [6, 7].

Comorbidity (the presence of other conditions in addition to a specified index condition) [8] may influence the effectiveness of treatments for specific conditions through competing risks, drug-drug, drug-disease and disease-disease interactions, altering the balance of risks and benefits [9‐11]. Underrepresentation of people with multimorbidity in clinical trials is therefore concerning.

However, most studies examining clinical trial representativeness have done so by analysing routine clinical practice data (e.g. from disease registers and electronic health records) to which trial eligibility criteria have been applied [12‐17]. Since factors other than eligibility criteria are likely to influence which people are recruited to clinical trials [18], such approaches provide only indirect evidence about the prevalence of comorbidity and multimorbidity in trial participants.

We examined the prevalence of comorbidity and multimorbidity among 122,969 participants from 116 industry-funded trials of novel drug therapies for 22 index conditions and compared these results with comorbidity and multimorbidity prevalence in 2.3 million patients living in the community.

Of the 124 trials meeting our inclusion criteria and made available via the CSDR and YODA repositories, 116 (including 122,969 participants for 22 index conditions) provided concomitant medication data allowing us to identify comorbidities. We had initially planned to include trials from the NIH BioLINCC repository, but found that none of the 8 trials which met our eligibility criteria provided sufficient data on comorbidities to be included in the analysis (Fig. 1). Index conditions are summarised in Table 1. Additional file 6 contains a summary of the characteristics of each trial. Additional file 7 shows summary statistics of the community sample for each index condition. Trials included in this analysis and trials which met our eligibility criteria but were not included (either because we did not obtain IPD or because the data we needed to perform these analyses had been redacted) were broadly similar in terms of the trial start dates, study design, excluding conditions and the number of participants enrolled as well as the clinical indications and drug classes studied (Additional file 8). However, we found that trials for inflammatory bowel disease and rheumatoid arthritis, as well as trials of immunosuppressant drugs, were somewhat overrepresented. We also found that while 11.3% of the IPD trials were phase 4 trials, 20.9% of non-IPD trials were phase 4, and that a lower proportion of IPD trials than non-IPD trials were very large (Additional file 8: Figure S8.1).

For each index condition, most comorbidities were more common in community patients than in the trials (Fig. 2). In community patients, the seven commonest comorbidities, from most to least common, were chronic pain, cardiovascular disease, arthritis, affective disorders, acid-related disorders, asthma or COPD, and diabetes. These conditions were common across all index conditions, although the ordering varied somewhat. For example, cardiovascular disease was commoner than chronic pain for both type 2 diabetes and COPD. This difference in ordering was evident for both the community sample and the trials. Indeed, for most index conditions, those comorbidities which were commonest in the community were also commonest for the trials.

For each of the comorbidities assessed, prevalence varied between trials. Some trials had a prevalence close to that of the primary care population, while in other trials the prevalence was much lower (Fig. 2 and Additional file 9). This pattern was similar across all index conditions, and for all comorbidities assessed. No specific comorbidities stood out as being consistently underrepresented. Conversely, none was found to be well represented across all trials.

Figure 3 shows the distribution of the comorbidity counts for trial participants and community-based patients. For each index condition, the comorbidity distribution for community-based patients lay to the right of the trial distribution (i.e. more comorbidities in community patients compared to trial participants). The community-based counts were standardised to the age-sex distributions of the trial participants for the relevant condition. However, the standardisation made little difference to the estimates (Additional file 10) so only the age-sex standardised results are presented. For the trial participants, where there were multiple trials per condition, the proportions were obtained from the modelled mean comorbidity counts for each index condition (see Table 2), under the assumption that the proportion of trial participants with each comorbidity count follows a Poisson distribution. Where there was only a single trial for a given condition (e.g. osteoarthritis), raw proportions are shown (see Additional file 5 for details). Comorbidity counts varied by index condition. Lower counts were evident for conditions such as asthma, inflammatory bowel disease and psoriasis. Conditions with higher comorbidity counts were those with a later age of onset. For most index conditions, the mean comorbidity counts were between 1.5-fold higher and 3-fold higher for community-based patients than for trial participants (Table 2).

Nonetheless, in absolute terms, comorbidity was common in both settings (Table 2). Most community-based patients had two or more comorbidities (i.e. three or more conditions overall) and would therefore be considered to have a high degree of multimorbidity under many definitions [28]. In trials, a significant proportion also had two or more comorbidities. This ranged from 10 to 15% for conditions such as asthma and psoriasis to around 40–60% for conditions with an older age of onset such as osteoporosis, dementia and pulmonary fibrosis.

On examining individual trials, the mean comorbidity count was the same or higher in the community than for every trial (Fig. 4). Nonetheless, there was considerable variation, even within the same index conditions. For some trials, the mean comorbidity counts were almost the same as in the community; for others, there was more than a twofold difference. In additional analyses, to explore this variation, we plotted the mean comorbidity count for each trial against trial-level characteristics such as the start date, phase, sponsor and total number of excluding conditions within the eligibility criteria, without observing any associations (Additional file 11).

We examined comorbidity and multimorbidity using individual-level participant data from 116 trials (122,969 participants) from seven industry sponsors of novel drug treatments for 22 common index conditions. We assessed the same comorbidities for the same index conditions from a nationally representative community sample of 2.3 million people. Comorbidity and multimorbidity were consistently lower in trial populations than in community patients, but were nonetheless common in both.

Our estimates of comorbidity in the community are consistent with previous findings: comorbidity was common, and for some conditions (e.g. COPD and osteoporosis), it was almost ubiquitous [1, 28]. To our knowledge, however, ours is the first study to compare comorbidity and multimorbidity patterns in the community to those in clinical trial populations by directly analysing comorbidity counts using individual-level trial participant data. In so doing, we confirmed that the mean comorbidity count for trials was approximately half that observed in the community.

We also found that, although patients with comorbidity or multimorbidity were underrepresented in many trials, comorbidity and multimorbidity were nonetheless common. For around half of the index conditions, the proportion of trial participants with ≥ 2 comorbidities (i.e. with three conditions and therefore highly multimorbid [28]) was above 30%. Given the ubiquity of multimorbidity among patients in the community [1, 28], it is perhaps unsurprising that comorbidity and multimorbidity are so common in industry-funded trials of novel drugs. However, we do not think that this unexpectedly high prevalence has previously been noted.

This finding is important because of current uncertainty as to the treatment of people with multimorbidity. Guidelines on the treatment of multimorbidity express reservations about the applicability of trial evidence to people with multimorbidity [29]. Moreover, in clinical practice, people with comorbidity (who, by definition, have multimorbidity) are less likely to receive certain drug treatments recommended across a range of disease-specific guidelines [2‐5]. Our findings that comorbidity and multimorbidity are underrepresented in clinical trials would support a cautious approach by guideline developers to the routine extrapolation of evidence. However, the finding that comorbidity and multimorbidity are common in clinical trials is important, because it suggests that trial data could potentially provide an important resource to allow treatment effects to be estimated in people with multimorbidity. These findings have implications for both triallists and for guideline developers.

The first implication for triallists and guideline developers relates to making better use of existing evidence. One way of doing so is via individual participant-level data meta-analyses. For this reason, we agree with the Alltrials initiative, and others, that sharing of IPD from clinical trials is crucial. Such analyses have helped resolve previous controversies about the efficacy of drugs in different sub-groups, showing, for example, that aspirin is similarly efficacious in men and women [30‐34]. Similar analyses have the potential to resolve similar controversies concerning comorbidity and multimorbidity [29, 35], potentially changing clinical practice, either by providing reassurance that trial findings can be applied to people with multimorbidity or by providing robust evidence to the contrary.

However, compared to meta-analysis of published results, IPD meta-analysis is costly and challenging. If trials are to be widely used to inform clinicians and guideline developers as to the efficacy of different treatments in the presence of comorbidity or multimorbidity, trials must publish results according to comorbidity sub-groups. Doing so will be challenging, however, because there are multiple different potential patterns of comorbidity. This is true even if only a small number of comorbid diseases are considered. There are, for example, 64 different possible ways that six conditions can occur together. Whether important and clinically relevant patterns of comorbidity can be identified from among such combinations remains an active and unresolved research question [36]. Nonetheless, we found that those comorbidities which were common in the community were also common in trials. Consequently, if clinically meaningful patterns of comorbidity and multimorbidity can be identified among people in the community, it may be possible to identify similar sub-groups among trial participants.

In the absence of consensus on which patterns of comorbidity should be grouped together, we propose that trials report treatment effects according to the presence/absence of common comorbidities, as well as by multimorbidity counts. Ideally, comorbidities would be defined using medical history data collected in a systematic and standardised manner across trials. In the absence of standardised medical histories [9, 37], some insights may be obtained from existing trials using drug-defined comorbidities, particularly where the focus is on conditions closely associated with particular drug classes (e.g. diabetes and glucose-lowering drugs) or on overall measures of multimorbidity, such as a count.

Despite these challenges, using clinical trial data to estimate treatment effects in people with comorbidity or multimorbidity remains appealing because of limitations in the alternatives. For example, observational datasets rich in multimorbidity, such as electronic health records, are used to estimate treatment effects. However, despite methodological advances in this use of observational data, it remains controversial, as unmeasured confounding can result in apparent treatment benefits when none really exist [9, 38].

The second implication for triallists relates to eligibility criteria and recruitment. For many indications, there was little difference in comorbidity counts between some trials and the community sample, whereas for other trials within the same indication the differences were large. This suggests that, even for standard industry-funded phase 3/4 trials, increasing the recruitment of comorbid participants is feasible. There is therefore potential for future trials to become more representative in terms of multimorbidity. In exploratory analyses, the differences in comorbidity between trials for similar indications were not related to start date, phase, sponsor or total number of exclusion criteria. Additional work is needed to identify the selection processes driving inclusion or exclusion of people with comorbidity so that trials can be made more representative. In addition, it will be important for future research to examine how conditions cluster in people with multimorbidity and whether this differs between clinical trial participants and people in the community in order to improve analysis and reporting of treatment effects as well as trial design.

The strengths of our study include large numbers and that the comorbidity definitions and analyses were pre-specified before making comparisons. However, there are several limitations. First, the trials collected medical history data in a variety of incommensurable ways. Consequently, we used concomitant medications to define comorbidities. This meant that some important conditions that are not treated with specific medications (e.g. chronic kidney disease) could not be identified reliably, whereas some other conditions which share treatments (e.g. asthma and COPD) had to be combined into broader categories. The use of some medications was so heterogenous as to preclude meaningful categorisation, and we did not attempt to use such drugs in any definition (for example, since amitriptyline is widely used in the treatment of chronic pain [39], we did not include it in our definition of affective disorders). Despite these limitations, some conditions are well defined by medications, and importantly, the same definitions were applied across trial and community data. Our community sample was taken from Wales because, while being broadly similar to the rest of the UK, it provides access to electronic medical records from a large and representative sample covering 70% of the population [40]. The Welsh population is broadly similar to the UK population in demography, and the findings are likely to be applicable to other high-income countries, but do require replication in other contexts. In order to facilitate this, we provide standard comorbidity definitions as well as data on the distribution of comorbidity counts, age and sex at the level of individual trials. A further limitation is that the included trials were not a random sample of all trials for these index conditions. Not all sponsors share trial data. Those who do share data do not make all trials available. Differences between trials that do or do not provide IPD may be a potential source of bias [41]. As such, we believe that the sharing of data by trial sponsors is to be encouraged, so as to minimise bias arising from the availability of a limited set of trials. Nonetheless, the included trials were similar to a wider body of registered trials across a range of characteristics (Additional file 8).

Clinical trial populations have a lower prevalence of comorbidity and multimorbidity than unselected community populations. Clinicians should exercise caution when applying disease-specific evidence and guidelines to people with comorbidity or multimorbidity. Nonetheless, comorbidity and multimorbidity are common in clinical trials. Given the limitations of observational data for estimating treatment effects, this suggests that standard industry-funded clinical trials are an underused resource for estimating treatment effects in multimorbidity. We would recommend that future disease-specific guidelines need to incorporate information concerning likely treatment effects in the context of the specific index condition and comorbidity or multimorbidity. To enable guideline developers to do so, triallists should at least report the prevalence of multimorbidity and a range of comorbidities among trial participants and should consider reporting treatment effect estimates stratified by comorbidity and/or multimorbidity. More general multimorbidity guidelines could also usefully include information in relation to this within any future guideline to permit more specific guidance for clinicians dealing with people with multimorbidity.