Predicting stress in automobile drivers using LSTMs with feature engineering

This study investigates the impact of stress on drivers and proposes a deep learning-based approach for predicting stress levels in near real-time. We utilized the publicly available DRIVE dataset, segmenting raw physiological signals (ECG, EMG, GSR, and RESP) into 10-second windows to capture short-term variations in stress indicators. We extracted novel time- and frequency- domain features and applied dimensionality reduction techniques including Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Random Forests (RF), and Autoencoders (AE) to identify the most informative attributes. Five resulting feature sets were evaluated with Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM) networks. Experimental results show that LSTMs consistently outperform the other architectures, achieving a maximum accuracy of 75.1% with a low standard deviation, while RF-based feature selection offers a practical trade-off between accuracy and sensor requirements. These findings highlight the potential of combining short time segments, advanced feature selection, and sequential deep learning models for efficient, real-time stress detection in driving scenarios.