Explainable AI framework for improved Thalassemia mental health classification and feature selection

Abstract

Mental health challenges in Thalassemia patients are often overlooked, despite their significant impact on quality of life. Traditional statistical and machine learning approaches often fail to capture the complex, nonlinear relationships between psychosocial and clinical variables, limiting both accuracy and interpretability. To overcome this, we propose the Adaptive Multi-Stage Ensemble with Dynamic Feature Interaction (AMSE-DFI)—a novel feature selection framework that dynamically integrates mutual information, ensemble learning, and graph attention mechanisms to capture intricate feature interdependencies often missed by traditional approaches. Using the SF-36 health survey data from 356 Bangladeshi patients, AMSE-DFI effectively identified key predictors such as total SF score, role emotional, and physical health summary, which collectively reflect both physical and psychological well-being. The model outperformed conventional approaches, showing strong predictive reliability and robust generalization, with SMOTE effectively addressing class imbalance in the clinical data. Importantly, Local Interpretable Model-Agnostic Explanations (LIME) based explainable AI offered clear, interpretable insights into how key features affect individual patient outcomes, making the model more understandable and actionable for clinicians. This framework provides a practical, transparent tool to support early detection and personalized management of mental health challenges in Thalassemia care.