Development and Internal Validation of an Explainable Longitudinal Voice Biomarker System for Relapse Risk Prediction in Schizophrenia: Hybrid Simulation and Real Audio Prototype Study

This record contains the preprint, supplementary outputs, and supporting materials for a study on an explainable longitudinal voice biomarker system for relapse risk prediction in schizophrenia. The project integrates acoustic feature extraction, temporal feature engineering, machine learning model development, explainability methods, and deployment oriented prototype components including a FastAPI backend, Streamlit dashboard, monitoring utilities, and frontend interface. The main predictive pipeline was internally validated on a longitudinal synthetic cohort with temporally engineered features. Real recorded human voice samples from consenting non schizophrenia participants were additionally used to test prototype audio feature extraction feasibility. Model evaluation included comparison of XGBoost, random forest, and logistic regression, together with confidence intervals, statistical testing, ablation analysis, and local explanation outputs using LIME. This upload is intended to support transparency, reproducibility, and scholarly dissemination of the research workflow and manuscript materials.