Machine Learning Approaches for Mental Health Relapse Prediction: Toward Clinical Deployment

Abstract

To some extent, the study was successful in realising the goal of precision psychiatry using ML and multi-modal features for the prediction of relapse in psychosis, bipolar and depressive disorders, with AUC between 0.69 and 0.89 as evidenced in. However, based on the problems stated in the systematic review, it found that only 18% of the models in the systematic review have a methodology that could be rated acceptable for the use of external validation, workflow/operationalization, transparency and generalizability. This work proposes a novel TRIPOD-compliant pipeline that implements XGBoost, Random Forests and Deep Neural Networks using SHAP/LIME explainers for explainability of the model. It uses the FHIR standard to handle more than 500 features comprising EHR features, digital biomarkers, and NLP features. It aims to answer questions beyond just a high predictive power and look to, among other issues, address federated learning for site-independent generalizability, high-speed inference on edge devices (within 1s), and identify the patients at highest risk. This paper plans to extend on the existing state-of-the-art in the area of relapse prediction to validate the model on a phase III RCT including more than 1,500 patients and later perform validation on a multi-ethnic and independent dataset, followed by a potential deployment. The system also focuses on external validation, calibration, and ethical implementation, which are significant translational challenges in precision psychiatry. The data is a summary of pooled results and design assumptions from previously published, validated research with intended implementation in real-life cohorts.