Stress is a complex psychological and physiological state that exerts profound effects on human health, performance, and quality of life. With the growing incidence of stress-related disorders among students, employees, and healthcare populations, accurate and timely stress detection has emerged as a critical research priority. Traditional approaches, though valuable in controlled laboratory settings, are constrained by reliance on manual feature engineering, limited adaptability, and poor scalability in real-world applications. Recent advances in deep learning have transformed this field by enabling automated feature extraction and multimodal data integration, spanning physiological signals (EEG, ECG, GSR), behavioral modalities (facial expressions, speech, and text), and wearable sensor data. This review provides a comprehensive synthesis of existing studies on deep learning–based stress detection, systematically examining convolutional neural networks, recurrent networks, long short-term memory architectures, attention mechanisms, and hybrid models. Benchmark performance across widely used datasets such as WESAD, SWELL, DREAMER, and DEAP is critically compared using metrics including accuracy, precision, recall, F1-score, and latency. Beyond performance evaluation, this study highlights challenges in data scarcity, generalizability across populations, computational complexity, and ethical considerations related to privacy and bias. Finally, future research directions are outlined, emphasizing opportunities in realtime stress monitoring, multimodal fusion, transfer learning, and privacy preserving frameworks. This review aims to serve as a structured and authoritative reference for advancing deep learning applications in stress detection and mental health monitoring.
Deeksha et al. (Tue,) studied this question.