An explainable Artificial Intelligence approach for delineating sex-based profiles in severe asthma

Introduction: In severe asthma (SA), a sex-oriented diagnostic and therapeutic approach is lacking; lung function limited to FEV1 investigates large airways involvement in pulmonary disease, with no correlation with symptoms. Aims: We aimed to delineate male and female profiles based on a comprehensive set of different clinical and functional features in patients with SA, using explainable Artificial Intelligence (AI) techniques. Methods: We collected the following data from 111 patients (63 women) with SA: anthropometric, physiological, clinical (APC), peripheral inflammatory biomarkers (PIB), complete lung function data and FeNO concentration (Spirometry + FeNO); and FOT measures. The AI framework included preprocessing steps, a classification procedure using an Extreme Gradient Boosting Classifier (XGboost) model, and the estimation of the SHapley Additive exPlanations (SHAP) values. We trained, validated, and tested the framework through a 100-times-repeated stratified and nested validation procedure. We reported mean area under the receiver operating characteristic (ROC) curve (AUC) and the bootstrap interval at 90% (int90%). Furthermore, we computed mean SHAP values in the outer test sets to obtain the contribution of each of the feature category. Results: The mean ROC AUC in the test set was 0.72 (int90% (0.62-0.82)). FOT and Spirometry + FeNO features exhibited superior predictive capabilities in classifying males vs. females, compared to PIB and APC. Conclusions: Comprehensive pulmonary function evaluation delineated a sex-linked profile in SA, paving the way for specific diagnostic, therapeutic and monitoring pathways for each patient according to sex: a “precision medicine” view in SA.