Explainable AI Models for Predicting Major Complications in Thalassemia Patients

Abstract

Thalassemia is an inherited blood disorder in which hemoglobin is synthesized abnormally and requires lifelong blood transfusions and continuous clinical monitoring. Secondary complications such as diabetes and leg ulcers are common, significantly reducing quality of life and increasing healthcare costs. Early detection of these complications is crucial for early intervention. However, the black-box nature of most machine learning algorithms hinders their application in clinical practice. In this study, predictive models for diabetes and leg ulcer complications in thalassemia patients were developed using a clinical dataset with 357 real patient data including demographics, laboratory values, and treatment histories. The models tested were Random Forest, XGBoost, and a Neural Network implemented in PyTorch. The neural network performed best for both complications, with excellent F1-scores and AUC scores. To ensure interpretation, Explainable AI methods were used using SHAP values to identify important clinical risk factors, LIME to produce case-specific explanations, and adversarial reasoning to test operational “what-if” scenarios. The results identified hypogonadism, serum ferritin, hemoglobin, ALT and AST as the best predictors of diabetes and leg ulcers. Through the application of explainable high-performance neural networks, this study demonstrates the potential of Explainable AI systems in improving personalized risk stratification, risk identification, and clinical decision-making for thalassemia management.

Ähnliche Arbeiten