Interpretable Machine Learning Framework for Predicting Treatment Resistance in Psychiatric Disorders Using Synthetic Pharmacogenomic and Clinical Data

This study presents a comprehensive and interpretable machine learning pipeline for predicting treatment resistance in psychiatric disorders using synthetically generated, multimodal data. The simulated dataset integrates gene expression profiles and clinically relevant features, such as age of onset, BMI, comorbidities, treatment response scores, and disease progression markers, thereby mimicking real-world pharmacogenomic and clinical complexities. The primary objective is to assess the feasibility of using artificial intelligence for individualized treatment stratification in psychiatry, with a strong emphasis on transparency and clinical relevance. Three classifiers—Random Forest, Gradient Boosting, and Calibrated Support Vector Machine (SVM)—were trained and evaluated using a balanced dataset of 2000 synthetic patients. Rigorous model validation was performed using key metrics, including ROC-AUC, F1-score, and balanced accuracy. Random Forest achieved the highest ROC-AUC (0.80) and balanced accuracy (0.72), followed closely by Gradient Boosting. Calibrated SVM exhibited lower performance but added methodological diversity. Feature importance was assessed using both traditional methods and permutation-based analysis, consistently high-lighting GENE_1, GENE_6, Rapid_Cycling, and Lithium_Response as dominant predictors. Local model explainability was further enhanced using LIME, which provided individualized insights into each prediction. A full visual clinical report was generated for a sample patient, including gene expression summaries, predicted drug responses, and actionable recommendations. The pipeline demonstrates a reproducible and explainable framework for AI-driven clinical decision support in psychiatry. By bridging genetic markers, treatment outcomes, and machine learning interpretability, this study offers a template for precision psychiatry using realistic simulation models. The findings reinforce the value of integrating interpretability into ML models to promote trust and applicability in clinical practice.