Multi-modal predictive modeling of schizophrenia severity: Leveraging liver function indicators and cognitive scores with random forest and SVM

Schizophrenia is a complex neuropsychiatric disorder with cognitive deficits and systemic physiological disturbances, including emerging links to hepatic dysfunction via the gut-liver-brain axis. Despite growing evidence, the integration of liver function biomarkers into predictive models for schizophrenia severity remains largely unexplored. This study proposes a multimodal machine learning framework combining hepatic indicators—Alanine Aminotransferase (ALT), Aspartate Aminotransferase (AST), Bilirubin, Albumin, and International Normalized Ratio (INR)—with cognitive assessment scores to enhance severity prediction. A synthetic dataset of 500 patient profiles was programmatically generated using MATLAB R2023a, simulating realistic clinical variability across demographics and biomarker distributions. Controlled missingness was introduced and imputed using moving mean methods, followed by Min-Max normalization to standardize features. Two machine learning models were developed: Random Forest for continuous severity regression and Support Vector Machine (SVM) with Error-Correcting Output Codes (ECOC) and Radial Basis Function (RBF) kernel for multiclass classification. The Random Forest regressor achieved an RMSE of 21.85, Mean Absolute Error of 17.26, and an R² of 0.70, capturing 72 % of variance. The SVM classifier attained 86.4 % accuracy, with macro-averaged precision, recall, and F1-score of 0.86, and an AUC of 0.91. Feature importance analysis revealed cognitive score, ALT, and AST as dominant predictors. Residual and confusion matrix analyses further confirmed model reliability. This integrative approach demonstrates the technical feasibility and clinical relevance of leveraging hepatic biomarkers alongside cognitive scores for schizophrenia severity assessment, offering a robust data-driven methodology for complex psychiatric evaluation.