Enhancing convolutional neural networks in electroencephalogram driver drowsiness detection using human inspired optimizers

Driver drowsiness is a significant safety concern, contributing to numerous traffic accidents. To address this issue, researchers have explored electroencephalogram (EEG)-based detection systems. Due to the high-dimensional nature of EEG signals and the subtle temporal patterns of drowsiness, there is increasing recognition of the need for deep neural networks (DNNs) to capture the dynamics of drowsy driving better. Meanwhile, optimizing DNNs architectures remains a challenge, as training these models is an NP-hard problem. Meta-heuristic algorithms offer an alternative to traditional gradient-based optimizers for improving DNNs performance. This study investigates the use of two human-inspired algorithms-teaching learning-based optimization (TLBO) and student psychology-based optimization (SPBO)-to optimize convolutional neural networks (CNNs) for EEG-based drowsiness detection. Results demonstrate strong predictive performance for both CNN-TLBO and CNN-SPBO, with area under the curve values of 0.926 and 0.920, respectively. TLBO produced a simpler model with 4,145 parameters, whereas SPBO generated a more complex architecture with 264,065 parameters but completed optimization faster (116 vs. 148 min). Despite minor overfitting, SPBO's efficiency makes it a cost-effective solution. In general, our findings contribute to the advancement of driver monitoring systems and road safety while emphasizing the broader role of meta-heuristic techniques in deep learning optimization.