A Multimodal Machine-Learning Framework for Predicting Lithium-Associated Renal Dysfunction from Synthetic Biomarker Trajectories

Long-term lithium therapy remains the most effective maintenance treatment for bipolar disorder, yet it poses a significant risk of progressive renal impairment in a subset of patients. Early recognition of individuals vulnerable to lithium-associated nephrotoxicity is clinically critical but challenging, given the heterogeneous evolution of renal biomarkers, nonlinear exposure effects, and complex interactions with comorbidities and concomitant medications. In this work, we introduce a multimodal machine-learning framework built on a mathematically transparent generator of synthetic renal tra-jectories. The simulator parameterizes baseline kidney function, lithium burden, hydration status, and comorbidity load to produce physiologically coherent serum and urine biomarker patterns over time. Using these synthetic trajectories, we train an ensemble of classical and gradient-boosted classifiers that achieve near-perfect discrimination of nephrotoxicity (AUC ≈ 1.00) on held-out data. Across models, biologically interpretable predictors including creatinine elevation, eGFR decline, and cumulative lithium load emerge consistently as dominant risk factors. Subgroup analyses confirm stable performance across age, sex, BMI, and serum-lithium strata, while SHAP explainability reveals mechanistic relationships linking lithium exposure, filtration dynamics, and acute renal stress. Together, these results demonstrate how structured synthetic data and interpretable multimodal machine learning can form a high-fidelity experimental testbed for studying lithium’s renal effects and for prototyping clinically actionable early-detection tools.