MHRD: Dual-Model Mental Health Risk Detection and Treatment Prediction System with Temporal Validation and Explainable AI.

The sudden transition to remote work during theCOVID-19 pandemic introduced severe systemic disruptions,raising critical questions regarding concept drift in occupationalmental health models. In this study, we propose the MentalHealth Risk Detection (MHRD) system, a dual-model machinelearning architecture designed to predict both workplace mentalhealth interference (Risk Model) and the likelihood of anemployee seeking professional help (Treatment Model). Utilizinga multi-year longitudinal dataset sourced from Open SourcingMental Illness (OSMI) spanning 2014 to 2021, we developeda gradient-boosting architecture utilizing LightGBM classifiersoptimized through a leakage-free SMOTE pipeline and MICEimputation. To rigorously test for concept drift, the Risk Modelwas subjected to strict temporal out-of-time validation (trained onpre-pandemic data, evaluated on pandemic data). The temporalmodel maintained a robust predictive accuracy of 61.70% (AUC:0.652), demonstrating Concept Amplification rather than totalconcept drift, and successfully outperformed standard baselineclassifiers. Simultaneously, the Treatment Model achieved 80.03%accuracy using a stratified longitudinal split. We integratedSHapley Additive exPlanations (SHAP) to provide explainableAI (XAI) insights and successfully deployed the framework asa real-time, FastAPI-driven cloud dashboard featuring dynamictemporal variance tracking, confidence analytics, and local featureimpact visualizations. This research provides a mathematicallyrobust, temporally validated decision-support system foroccupational health analytics.