A Hybrid Ensemble Data Mining Framework Integrating Multimodal Behavioral, Physiological, and Social Media Signals for Early Detection and Risk Stratification of Mental Health Disorders

Abstract

Mental health disorders, including depression, anxiety, bipolar disorder, and stress-related conditions, represent one of the leading causes of global disease burden, yet a large proportion of cases remain undiagnosed until symptoms become severe. Conventional screening relies on retrospective self-report instruments and clinician judgement, which are inherently subjective, episodic, and unable to capture the continuous, multidimensional signatures of psychiatric risk as they emerge in daily life. This paper proposes a Hybrid Ensemble Data Mining Framework (HEDMF) that integrates three complementary modalities — behavioral data (smartphone usage, keystroke dynamics, sleep and activity patterns), physiological data (heart rate variability, electrodermal activity, sleep architecture from wearable sensors), and social media data (linguistic, temporal, and network-level markers of affect) — into a single computational pipeline for early detection and risk stratification of mental health disorders. The framework combines modality-specific preprocessing and feature engineering with a hybrid fusion strategy that merges early (feature-level) and late (decision-level) fusion through a stacked ensemble architecture comprising Random Forest, Gradient Boosting (XGBoost/LightGBM), Support Vector Machines, and a Bidirectional Long Short-Term Memory (BiLSTM) network as base learners, combined via a meta-learner and calibrated probability outputs. A downstream risk-stratification module converts ensemble outputs into ordinal risk tiers (low, moderate, high, urgent) using a cost-sensitive thresholding scheme intended to support triage rather than diagnosis. An explainability layer based on SHapley Additive exPlanations (SHAP) and attention-weight visualization is incorporated to expose the behavioral, physiological, and linguistic markers driving each prediction, supporting clinician trust and actionable interpretation. The framework is evaluated on a composite, ethically sourced multimodal dataset assembled from open behavioral, physiological, and social-media benchmark corpora, supplemented by a synthetic cohort generated to preserve realistic feature distributions while protecting privacy. Experimental results, reported here as an illustrative proof-of-concept case study, indicate that the hybrid ensemble achieves higher accuracy, sensitivity, and F1-score than single-modality baselines and than early-fusion-only or late-fusion-only ensembles, while the risk-stratification module demonstrates favorable calibration on the high-risk tier where false negatives carry the greatest clinical cost. The paper further discusses ethical, privacy, and deployment considerations relevant to consent-based passive sensing and social media analysis in mental health contexts, and outlines a roadmap for prospective clinical validation. The proposed HEDMF contributes a modular, extensible, and explainable architecture that can serve as a reference design for future multimodal digital-phenotyping systems aimed at prevention-oriented mental health screening.