EEG quantization and entropy of multi-step transition probabilities for driver drowsiness detection via LSTM

Detecting driver drowsiness through electroencephalogram (EEG) poses challenges due to the complexity and variability of brain activity across different subjects. This study proposes a feature extraction pipeline combined with a Long Short-Term Memory (LSTM) network. EEG data from each electrode channel is normalized and quantized to discrete levels, where the probabilities of transitioning to other levels in the subsequent timesteps is modeled using multiple Hidden Markov Models (HMMs) for different timestep shifts. From the HMM emission probabilities, Shannon, Renyi, Tsallis, and Min entropy are extracted, forming a feature set that captures temporal channel information. These features are input into an LSTM network to classify alert or drowsy states. Monopolar and bipolar EEG montages are also investigated. Experiments on balanced and unbalanced EEG datasets show higher performance compared to existing machine learning and state-of-the-art deep learning methods. Subject-wise 5-fold and leave-one-out cross-validation achieved 91.23 % and 81.88 % accuracy on the balanced dataset, and 91.38 % and 80.58 % accuracy on the unbalanced dataset. Saliency analysis and ablation studies identifies key EEG channels, time-step shifts, and quantization levels contributing to drowsiness detection. The Oz channel produces the most discriminative features and offers state-of-the-art single-channel accuracy of 74.66 %. The feature extraction algorithm provides significant distinction between drowsy and alert classes, with transition Shannon entropy value being higher for the drowsy class. Reasons behind confidently correct and incorrect predictions and model criteria for detecting drowsiness are also explored. This work highlights the potential of using signal quantization with entropy of transition probabilities to extract meaningful features.