Machine learning pinpoints response patterns with mepolizumab in severe asthma
Predicting who benefits most from mepolizumab has been one of the practical problems in severe asthma care. The anti-IL-5 biologic reliably reduces exacerbations in many patients, yet real-world response remains uneven, and that variability complicates treatment selection when clinicians are trying to spare patients from repeated systemic corticosteroid exposure.
A machine learning analysis of the REALITI-A cohort identified a small set of clinical variables that repeatedly tracked with response across three separate definitions of benefit: change in clinically significant exacerbations at Year 1, being free from clinically significant exacerbations at Year 1, and maintaining that response at both Year 1 and Year 2.
Exacerbation burden and symptom control carried the strongest signal
The primary and secondary analyses included 685 patients, while the exploratory maintained-response analysis included 529 patients. Across those models, the most consistent pattern was straightforward: a heavier exacerbation burden before treatment and poorer asthma control metrics tended to separate responders from non-responders. In the primary analysis, the XGBoost model selected higher pre-exposure clinically significant exacerbation rate as a top predictor, alongside lower body mass index, lower total immunoglobulin E, lower baseline activity impairment, and lower baseline Asthma Control Questionnaire-5 score.
Those features are clinically plausible. Patients with more frequent exacerbations before starting biologic therapy often have a more active inflammatory phenotype, and a therapy targeting interleukin-5 would be expected to have the clearest effect where eosinophilic inflammation is driving instability. Baseline control scores and activity limitation also matter because they reflect how much disease burden remains before the first dose even reaches steady effect.
One detail stands out sharply: lower baseline asthma burden was not the only marker associated with response. Several of the top predictors pointed toward patients who were already symptomatic but not necessarily the sickest in every dimension, suggesting that response to mepolizumab may depend on a narrower inflammatory and functional profile than simple severity labels capture.
Different models, overlapping predictors
For the secondary endpoint, the random forest model identified lower pre-exposure clinically significant exacerbation rate, residency outside the United Kingdom, and lower baseline Asthma Control Questionnaire-5 score, body mass index, and baseline activity impairment as the top predictors. In the exploratory analysis, XGBoost again placed residency outside the United Kingdom among the top predictors, together with lower pre-exposure clinically significant exacerbation rate, lower baseline Asthma Control Questionnaire-5 score, eosinophil levels, and body mass index.
The repeated appearance of Asthma Control Questionnaire-5 score, body mass index, and exacerbation history across models suggests these variables carried stable predictive weight, even as the endpoint changed. Residual symptom burden before treatment appears to be more than a background descriptor; it behaves like a measurable clinical fingerprint of how a patient might respond once eosinophilic inflammation is suppressed.
Residency outside the United Kingdom also emerged as a predictor in both the secondary and exploratory analyses. The study design does not establish why geography entered the models, and no causal explanation can be drawn from these data alone. Differences in practice patterns, referral pathways, baseline disease characteristics, or follow-up structure could all have contributed, but the analysis stops short of identifying which of those possibilities, if any, drove the signal.
Why the biology fits the model outputs
Mepolizumab acts on the IL-5 pathway, reducing eosinophil-mediated airway inflammation. That mechanism explains why eosinophil levels, prior exacerbation frequency, and pre-treatment symptom control would surface as informative predictors in a real-world dataset. In practical terms, the drug is less likely to be a universal solution than a targeted tool, and machine learning can help expose the combinations of variables that mark a patient as more likely to benefit.
The analysis also suggests that response is not defined by a single laboratory value. Eosinophils mattered, but so did body mass index, asthma control, and activity impairment. In other words, the treatment signal was embedded in a composite clinical state rather than one biomarker alone.
Real-world prediction still depends on the quality of the input data
The study drew from the prospective REALITI-A cohort, which enrolled 822 patients across seven countries between December 2016 and October 2019. Data for the current analysis were deidentified and accessed later for secondary evaluation. Thirty clinically relevant features were expanded into 122 variables, including the top medications and comorbidities among eligible patients, allowing the algorithms to examine a broad clinical field rather than a narrow biomarker panel.
Investigators compared linear and logistic regression, random forest, and XGBoost models, then used the best-performing model for feature selection. A SHapley Additive Explanations post-hoc analysis was also performed to interpret the rankings. Those methods strengthen the internal logic of the findings, but they do not remove the limits of the source data: observational cohorts cannot fully separate treatment effect from the clinical context in which treatment was chosen.
The cohort exclusions also matter operationally. Patients missing exacerbation data or lost before Year 1 or Year 2 post-exposure were excluded from the corresponding analyses, reducing the available sample for some endpoints. The predictive patterns are therefore anchored in patients with enough follow-up and recorded data to support modeling, not in the full breadth of all real-world mepolizumab recipients.
What clinicians can use now
For severe asthma care, the practical message is not that an algorithm should replace judgment. It is that a patient with frequent pre-treatment exacerbations, lower baseline asthma control, lower activity impairment, and an eosinophilic profile may fit the response pattern most consistently associated with mepolizumab benefit in this dataset.
The work also points to a broader shift in biologic prescribing. Severe asthma is not a single disease state, and treatment response is increasingly likely to depend on phenotype, symptom burden, and comorbidity pattern rather than diagnosis alone. A model that combines those variables may eventually support more precise biologic selection at the bedside, especially when clinicians are trying to decide whether to continue, switch, or escalate therapy after the first year.
What remains unresolved is the degree to which these predictors will generalize outside the REALITI-A setting and whether similar models can be converted into usable decision tools in routine practice without losing accuracy. The cohort provides a strong starting point, but the final test will be whether the same predictors hold in independent severe asthma populations treated under different health systems and follow-up structures.