MULTIMODAL STRESS MARKER RECOGNITION USING ARTIFICIAL NEURAL NETWORKS: TOWARD REAL-TIME, PERSONALIZED STRESS ASSESSMENT

Stress is a complex psychophysiological condition that affects well-being and long-term health. Objective stress detection through physiological signals holds great promise for clinical, occupational, and sports medicine applications, enabling adaptive, real-time mental state assessment. This study develops a multimodal artificial neural-network (ANN) framework for stress marker recognition using electroencephalography (EEG), electrocardiography (ECG) and validated self-report data. 29 participants completed 2 randomized sessions - control and stress - each comprising Stroop, mental-arithmetic, and Tetris tasks. Stress conditions introduced time pressure, negative feedback, and visuomotor perturbations. Signals were recorded from 32 EEG channels together with ECG and pre-/post-session STAI and PANAS questionnaires. The ANN, implemented in PyTorch, integrates dual EEG branches (temporal-convolutional and spatial-CNN), GRU layers for ECG and respiration, and an attention-based multimodal fusion module to dynamically weight input importance. Physiological validation confirmed effective stress induction: heart-rate variability (RMSSD) decreased significantly under stress, indicating reduced parasympathetic tone. EEG topographic analysis showed increased left-lateralized frontal delta and frontal low-gamma power (~30 %), and reduced fronto-central beta, consistent with a shift from synchronized low-frequency to desynchronized high-frequency cortical activity. Importantly, frontal asymmetry analyses revealed left-lateralized activation in low-anxiety groups, whereas low-reactivity individuals exhibited more symmetric delta and gamma responses. Group-based contrasts (high/low anxiety × high/low reactivity) highlighted distinct neurophysiological stress signatures across personality profiles. Questionnaire scores corroborated elevated negative affect and anxiety under stress. Preliminary ANN training achieved a reasonable accuracy distinguishing stress from control states. The framework demonstrates promise for personalized, real-time stress assessment and future integration into wearable or digital-health platforms.