SmartDrowsyNet: A TCN Deep Learning Framework for early EEG-Based Drowsiness Detection

Abstract

Driver drowsiness is a major contributing factor to road accidents worldwide, making its early detection a critical challenge for intelligent transportation and driver assistance systems. Most existing approaches rely on observable indicators such as eye closure, head movements, facial expressions, or steering behaviour. Although these methods can identify fatigue-related symptoms, they generally react only after the driver has already become impaired and provide limited insight into the underlying cognitive state. In contrast, electroencephalography (EEG) provides direct access to brain activity, enabling the detection of early neural signatures of drowsiness before visible behavioural symptoms emerge.

In this work, we investigate EEG-based drowsiness estimation using the SEED-VIG dataset, which contains EEG recordings from 23 subjects performing a two-hour simulated driving task designed to induce fatigue. Rather than formulating drowsiness detection as a binary classification problem, we address it as a regression task aimed at predicting a continuous vigilance indicator, namely the Percentage of Eye Closure (PERCLOS), which quantifies the proportion of time during which the driver's eyes remain closed.

To capture temporal dependencies in EEG dynamics, we propose SmartDrowsyNet, a lightweight deep learning architecture that combines a Temporal Convolutional Network (TCN) with a lightweight temporal attention mechanism. The model operates on Differential Entropy (DE) features extracted from the five conventional EEG frequency bands (delta, theta, alpha, beta, and gamma), complemented by physiologically motivated inter-band ratio features. The TCN captures long-range temporal dependencies, while the attention layer selectively emphasizes the most informative temporal patterns for vigilance estimation.

Experimental results demonstrate that the proposed framework consistently outperforms a baseline Long Short-Term Memory (LSTM) model in continuous PERCLOS prediction. The proposed architecture achieves accurate tracking of drowsiness evolution across subjects and reaches a state-classification accuracy of 71.8% when vigilance levels are grouped into three categories: Alert, Transitional, and Drowsy. Furthermore, the model exhibits strong temporal modelling capabilities while maintaining a lightweight computational footprint suitable for real-time deployment.

The proposed framework provides a complete and reproducible pipeline for EEG-based vigilance monitoring, including preprocessing, feature extraction, temporal window construction, model training, and evaluation. Its predictive performance and computational efficiency make it a promising candidate for future wearable and embedded driver-monitoring systems aimed at reducing fatigue-related road accidents.