We evaluated the combined capacity of salient patient characteristics in predicting the benefit of mepolizumab in reducing severe exacerbation rates and improving asthma symptom control. We quantified their capacity in terms of disparities in treatment response via the Lorenz curve and the Gini index, and in terms of how the observed treatment benefit changed over quintiles of predicted benefit. Patient characteristics captured higher heterogeneity in individual treatment response for improving asthma control than for exacerbation reduction (Gini indices of 0.35 vs. 0.24). Similarly, treatment benefit varied more widely across quintiles of predicted benefit for asthma control than for exacerbation rate. As well, patient characteristics showed greater potential for predicting below-average than above-average treatment response. On the other hand, while patient characteristics could identify subgroups with low treatment response, they did not identify any subgroup that would be harmed by treatment (while observed treatment benefit was negative in the bottom two quintiles for ACQ5 score, it did not achieve statistical significance).
How do these results inform the development of precision medicine for severe asthma? This was a feasibility study, not ready for dissemination as a clinical prediction model. Such an algorithm would require a more formal model development and external validation approach. However, before embarking on such a process, it is beneficial to know whether patient characteristics can indeed explain heterogeneity in treatment effect. Our results show that a precision medicine approach based on clinical prediction modeling could indeed result in more efficient treatment assignment rules, even among patients who have satisfied the strict eligibility criteria of severe asthma trials. This finding can have important implications. A recent systematic review concluded that the prices of biologics would need to be reduced by at least 60% to meet the threshold for cost-effectiveness [
37]. An alternative way to improve the value-for-money potential of biologics is to enable ‘concentration of benefit’ by adopting a stratified treatment algorithm that prioritizes treatment to individuals with high likelihood of treatment response. For instance, based on our results, a 60% improvement in the efficiency of mepolizumab would be achieved in terms of asthma symptom control if it were given to the 40% of eligible patients with the highest likelihood of improvement in symptom control, or to the 20% of eligible patients with the highest predicted reduction in exacerbations. These results were obtained despite the fact that patients were already pre-selected for the likelihood of response to biologics, as reflected in the inclusion criteria of the two RCTs. Our findings therefore indicate that a more accurate selection for treatment based on a combination of patient characteristics is feasible. Such gain in efficiency of treatment decisions is unlikely to be achieved by the use of a single biomarker. This was demonstrated by the significantly better performance of the combined set of patient characteristics over BEC alone, which is an important component of biologic eligibility criteria [
38]. Further, the finding that patient characteristics were more capable of identifying below-average compared with above-average response suggest a potentially high utility of prediction models for ruling out treatment, or otherwise identifying patients who might require more intense monitoring and evaluation due to lower likelihood of treatment response. Finally, the importance ranking of predictors from our study can inform the selection of predictors or the development of clinical prediction models for treatment response.
To the best of our knowledge, this study is the first to quantify the combined capacity of patient characteristics towards predicting treatment response to a biologic in patients with severe asthma. Use of standardized clinical measurements and robust statistical analyses enhance the validity of our findings. This study also has several limitations. First, the generalizability and applicability of the current findings require careful consideration because the clinical trial data were limited by the narrow eligibility criteria of the RCTs (already selecting patients for high likelihood of response) and relatively small sample size. In addition, exacerbation rate and asthma control in the comparison group might be biased estimates of their counterparts in the general severe asthma population due to a placebo effect of participating in a clinical trial [
39]. However, real-world prescribing practice often deviates from guidelines for use of biologic therapies, which may also introduce bias to the prediction affecting the generalizability of our findings. For instance, a US-based claims data analysis showed that the majority of asthma patients initiating biologic therapy did not have severe asthma and many were poorly adherent to first-line asthma treatment, indicating inappropriate treatment escalation [
40]. Therefore, external validation of the development of clinical prediction models should be performed in real-world data with adjustment for medication adherence and other real-world factors for biologic initiation [
41], which should include a broader class of biologic drugs. Further, our analysis was performed with only one biologic (i.e., mepolizumab) at a single intravenous dose (75 mg) which was available in both MENSA and DREAM, whereas the effect of biologics may vary across type, dosage, and patient population. The more widely used 100-mg subcutaneous mepolizumab was available only in MENSA which could not produce a sufficient sample size to perform the desired shrinkage [
25]. Nonetheless, improvements in asthma exacerbation and other outcomes such as FEV1 were similar between these two doses and modes of delivery in MENSA [
10]. Third, other domains of treatment response, such as reduction in OCS exposure and improvement in quality of life, were not explored in this analysis. Last but not least, other important risk factors such as FeNO level, comorbidities, atopy status, socioeconomic status, environmental exposures, and other medication use were not recorded in the data and thus their predictive capacity could not be explored. On the other hand, several prognostic factors such as biomarkers and lung function parameters are not routinely collected in clinical practice, particularly in primary care, which could limit the applicability of a full prediction model in the future.