Predicting real-world response to mepolizumab in severe asthma using machine learning

Mepolizumab reduces clinically significant exacerbations (CSE) in patients with severe asthma, but research is needed to better understand the factors driving response. This study aimed to identify predictors of mepolizumab response using machine learning. 122 variables from 685 patients of REALITI-A study were used as model inputs. Linear regression, random forest, and XGBoost were tested. The best performing algorithm based on R² was used to train the model to identify predictors of response, defined as CSE change between pre-exposure (1-year and run-in) and 1-year post-exposure. Shapley Additive Explanations (SHAP) ranked variables based on feature importance and estimated the direction of relationship between features and the outcome. The best performing algorithm was XGBoost (R²=0.83). The top 5 ranking predictors were: pre-exposure rates of CSE, body mass index (BMI), immunoglobulin E (IgE) levels, activity impairment, and baseline asthma control questionnaire (ACQ) score, with SHAP values of 2.25, 0.18, 0.12, 0.11 and 0.09, respectively (Figure 1). Furthermore, higher rates of pre-exposure CSE directly correlate, while higher BMI, activity impairment, and baseline ACQ scores inversely correlate with greater reductions in CSE. This study identified top predictors of response to mepolizumab and may facilitate developing a clinically relevant model to support improvement of outcomes for patients with severe asthma. <fig><object-id>erj;64/suppl_68/PA448/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig>