Abstract
Visceral Leishmaniasis (VL), also known as Kala-Azar, poses a significant global public health challenge and is a neglected disease, with relapses and treatment failures leading to increased morbidity and mortality. This study introduces an explainable machine learning approach to predict VL relapse and identify critical risk factors, thereby aiding patient monitoring and treatment strategies. Leveraging data from a follow-up study of 571 patients, the survival machine learning models are applied, including Random Survival Forest (RSF), Survival Support Vector Machine (SSVM), and eXtreme Gradient Boosting (XGBoost), for relapse prediction. The results demonstrated that RSF, with a C-index of 0.85, outperformed the conventional Cox Proportional Hazard (CPH) model (C-index 0.8), offering improved prediction capabilities by capturing non-linear relationships and variable interactions. To address the lack of transparency (in terms of feature importance) in Machine Learning (ML) models, the SHapley Additive exPlanation (SHAP) method is employed, which enhances model interpretability (feature importance) through visual insights. SHAP dependence plots allowed the healthcare professionals to evaluate which factors encourage the occurrence of the relapse. A statistically significant relationship between HIV co-infection (HR=3.92, 95% CI=2.03-7.58) and VL relapse was identified through -2 log-likelihood ratio and chi-square tests. These results indicate the promise of explainable artificial intelligence (XAI) for making clinical decisions and remedying recurrences in VL.