Comparing Classic Machine Learning with Deep Learning for Stress Detection Using Wearable Sensors

Based on conducted research, stress can have a significant impact on human relationships and human-related incidents. By identifying stress during daily activities such as driving, some incidents and accidents can be prevented. In this study, the PhysioNet database pertaining to drivers' heart rate during driving was utilized, and their features were extracted. Subsequently, the features underwent reduction using PCA and were compared using two artificial intelligence methods. The results, including accuracy, error, and validation credibility with fold-10 in four classes, were obtained for both neural network and deep learning approaches. In the feature extraction phase, 7 spatial features, 16 frequency features, and 64 wavelet features were employed. The classification result for the neural network achieved an accuracy of 90.8%±0.8. In the deep network, comprising one-dimensional CNN and Dense layers, with a fusion of raw signals and extracted features, the accuracy reached 96.3%±0.6. These findings indicate the superiority of deep learning over neural networks in this domain. This diagnostic system is suitable for portable and compact applications in individuals' daily activities.