%0 Journal Article
%T Lithium Associated Chronic Kidney Disease Prediction Using Explainable Machine Learning: A Comprehensive Modelling and Interpretation Framework
%A Rocco de Filippis
%A Abdullah Al Foysal
%J Open Access Library Journal
%V 13
%N 3
%P 1-36
%@ 2333-9721
%D 2026
%I Open Access Library
%R 10.4236/oalib.1114920
%X Lithium remains the most effective long-term treatment for bipolar disorder, yet its therapeutic benefits are offset by a well-established risk of chronic kidney disease (CKD). Anticipating lithium-associated renal impairment is clinically challenging because the underlying mechanisms are subtle, multi-variate, and evolve dynamically with cumulative exposure. In this study, we develop a transparent, end-to-end machine-learning framework for early detection of lithium-induced kidney damage using a comprehensive synthetic cohort that incorporates lithium duration, serum concentrations, renal function biomarkers, comorbidities, metabolic factors, and inflammatory markers. Four supervised classifiers, Random Forest, XGBoost, logistic regression, and CatBoost were evaluated using stratified 5-fold cross-validation. CatBoost achieved the strongest generalization performance (test AUC = 0.785) and was subsequently selected for full explainability analysis. To ensure clinical interpretability, we integrate a multi-layered explanation suite comprising SHAP-based global and local attributions, interaction effect quantification, patient-specific waterfall plots, LIME explanations, bootstrap con-fidence intervals, permutation-based statistical validation, and correlation analyses between feature values and model-derived risk. Across methods, lithium exposure metrics (duration, serum level, dose), renal decline markers (eGFR trajectory, creatinine, proteinuria), hypertension, and key interaction terms (duration กม level; age กม duration) consistently emerge as dominant predictors of early renal damage. The SHAP-derived risk landscape reveals coherent, monotonic associations between cumulative lithium exposure, deterioration in renal function, and elevated CKD risk, aligning closely with established nephrotoxic pathways. Although the dataset is synthetic, the modelling strategy provides a rigorous blueprint for interpretable AI driven nephrotoxicity surveillance. The framework offers actionable, clinician-ready insights and establishes a foundation for future validation on real-world lithium cohorts, supporting precision monitoring and early intervention in lithium-treated patients.<br />
%K Lithium Nephrotoxicity
%K Chronic Kidney Disease
%K Explainable AI
%K SHAP
%K CatBoost
%K Clinical Risk Modelling
%K Machine Learning
%U http://www.oalib.com/paper/6888125