Supporting Hemodialysis Decision-Making in Lithium Poisoning: An Explainable and Clinically Interpretable Machine Learning and Nomogram Development

Lithium poisoning (LP) poses a critical management challenge due to its narrow therapeutic index and unpredictable toxicodynamics. Present paper aims to use machine learning capabilities to develop a decision-support system for risk assessment of hemodialysis in LP. We analyzed 207 patients with LP admitted to a referral toxicology center, of whom 27 (13.0%) required hemodialysis. Following a feature selection strategy, four algorithms, logistic regression, support vector machine, artificial neural network, and random forest, were trained. with 5-fold cross-validation. The approach was focused on preventing data leakage into the validation and test. The random forest model outperformed other models with a test-set AUROC of 0.83, sensitivity of 80.0%, specificity of 83.0%, and F1-score of 0.53. Mean cross-validation AUROC was 0.90. SHAP analysis identified serum lithium level, neurological symptoms, alkaline phosphatase, and age as key predictors. Platt recalibration improved the Brier score from 0.126 to 0.086 and calibration slope to 1.06. Decision curve analysis showed net clinical benefit across a wide threshold range. Bedside nomogram increased clinical utility (AUROC = 0.79) by classifying patients into low- moderate- and high-risk for hemodialysis. Although this decision-support system can significantly help the clinicians, an external validation and future studies using a bigger development set is required.