Ai-driven multi-model stress detection using physiological dynamics

Stress is a complicated psychophysiological situation enormously impacting the health, quality of life, and productivity and is directly related to the development of multiple physical and mental illnesses. This study introduces a multimodal dataset-level stress classification framework for synchronised electrocardiogram (ECG), electromyogram(EMG), and blood volume pulse (BVP) signals based on the WESAD dataset, a recording of 15 subjects in a controlled laboratory environment. The suggested pipeline combines physiologically relevant preprocessing, adaptive segmentation, feature-level multimodal fusion, and machine-learning classification to model cardiovascular, muscular, and vascular stress responses. Over 60 million raw samples were processed to form 37,498 multimodal windows, which were cross-validated using a rigid Leave-One-Subject-Out (LOSO) protocol to ensure independent assessment of subjects. In this assessment, Logistic Regression demonstrates methodological feasibility under controlled dataset conditions. Compared to deep learning methods, classical machine learning models demonstrated competitive stability under LOSO conditions, though performance may vary with larger or more diverse datasets. The class imbalance issue (11.5% stress samples) was addressed using class weighting and majority-vote window labelling. The computer analysis has a low inference latency (0.183 ms) and a high throughput (5,464 windows/s), indicating offline feasibility under controlled dataset conditions, without implying readiness for real-time or hardware deployment. The framework offers interpretable physiological characteristics that can be used in studies of mental health in the workplace, remote stress measurement, and high-stress job studies. At the methodological level, the work is relevant to SDG 3 (Good Health and Well-Being) and SDG 8 (Decent Work and Economic Growth), as it provides a reproducible foundation for stress analysis studies.