Enhancing Stress Detection Accuracy: A CNN-LSTM Model for ECG Signal Analysis

Stress detection using physiological signals is crucial for modern healthcare and well-being systems, particularly given the increasing prevalence of stress-related disorders. A major challenge in multi-class stress classification using ECG is the significant overlap in physiological characteristics between different stress states, making clear differentiation difficult due to the high variability of human physiological responses. This study addresses this issue through the selection of the most informative ECG features and the development of a hybrid neural network architecture. The proposed CNN-LSTM architecture consists of two main blocks: a convolutional block with two Conv2D layers, and a recurrent block with two consecutive LSTM layers. This architecture optimally processes 14 carefully selected ECG parameters from the time-domain (7 features), frequency-domain (4 features), and nonlinear-domain (3 features) analyses, utilizing spatial and regular dropout mechanisms to prevent overfitting. Using the publicly available WESAD dataset, the model achieves an overall accuracy of 97% in classifying three states: no stress, amusement, and stress. Notably, the minimal error rate between no-stress and high-stress states (less than 1%) makes the system highly reliable for practical applications, providing an effective single-sensor approach for real-time stress monitoring.