Explainable Machine Learning Model for Personalized Assessment of the Need for Intensive Care in Patients with Myasthenia Gravis Based on Phenotype and Inflammatory Markers

Background Early identification of myasthenia gravis (MG) severe exacerbations that require intensive care unit (ICU) admission remains challenging due disease highly heterogeneous. Existing models seldom incorporate phenotype stratification, limiting interpretability and clinical utility. Aim This study aimed to develop an explainable machine learning (ML) model to assess phenotype-specific risk factors and predict ICU admission for patients with MG. Design and methods This retrospective study included 319 hospitalized patients with MG. We utilized clinical variables and systemic inflammation markers for ML modeling. Subgroups were prespecified by thymoma status and age at onset. Primary outcome was need for intensive care during the hospitalization. Model performance was assessed. SHapley Additive exPlanations (SHAP) was employed for interpretation of global and individual features. Results Of 319 admissions, 33 (10.3%) required ICU care. Four ML models were trained with nested 3-fold validation. eXtreme Gradient Boosting performed best (AUC: 0.94). SHAP ranked MGFA clinical classification, infection during admission, body mass index, and inflammatory indices as principal predictors. Subgroup analyses revealed phenotype-specific patterns, including opposite SHAP directions for inflammatory markers between thymoma and non-thymoma groups and distinct rankings between different age of onset groups. SHAP-based personalized explanations revealed patient-specific variations in risk ranking and directionality of predictive features. A web-based calculator was developed to support real-time clinical implementation of the model, potentially allowing for risk stratification. Conclusions An explainable machine learning framework that integrates clinical features with systemic inflammation indices reliably predicts ICU admission in MG and reveals phenotype-specific risk patterns, enabling personalized risk stratification in MG management.