"Mental Health Analysis (Stress Level Detection) Using Machine Learning Models" رسالة ماجستير

Abstract

This research establishes an intelligent framework for the early identification of stress and adaptive intervention via multimodal data integration and real-time decision-making. The research incorporated three diverse datasets: WESAD (Wearable Stress and Affect Detection Dataset- physiological signals), ExtraSensory (behavioral and environmental data), and CAMS (Cognitive & Affective Mind-States Dataset - psychological and cognitive perspective), a psychological and linguistic dataset developed specifically for this study. In the absence of a cohesive multimodal dataset, a unique state-based fusion methodology was developed, aligning all modalities based on stress levels (baseline, amusement, tension) rather than participant identity. This technique produced a unified multimodal dataset that integrates physiological, behavioral, and textual markers. A quantitative machine learning approach was employed, encompassing data cleansing, feature engineering, correlation analysis, Random Forest significance, and z-score standardization. The modeling pipeline included modality-specific logistic regression models, a stacking meta-learner for multimodal integration, probability calibration (Platt scaling, Isotonic regression), and a hybrid rule-based component to enhance early detection through dynamic thresholding. The system was additionally enhanced with a JITAI module V employing a contextual bandit algorithm to tailor therapies according to real-time physiological feedback. Results indicated that physiological signals were the most robust predictors of early stress (ROC-AUC = 0.983), whereas behavioral aspects exhibited little predictive capacity, and textual-emotional markers provided further insights. Feature-level multimodal fusion attained the maximum accuracy (ROC-AUC = 0.996). Dynamic thresholding markedly enhanced recall for high-risk conditions, reinforcing the purpose of early intervention. Among the intervention options, calm breathing yielded the greatest reward, succeeded by cognitive reappraisal. The integrated JITAI model exhibited ongoing learning and adaptation, delivering tailored and context-sensitive recommendations. The study advocates for the implementation of the suggested framework in digital mental health platforms, enhancing dataset variety, and performing human-centered assessments to confirm the ecological validity of early intervention systems.