ENHANCING DROWSINESS CLASSIFICATION MODEL PERFORMANCE THROUGH NEURAL NETWORK OPTIMIZATION

Abstract

  Drowsy driving is a significant cntributor to traffic accidents, prompting the need for effective drowsiness detection systems. This research introduces a novel approach, the Connectivity CNN-LSTM, for real-time drowsiness classification based on Electroencephalography (EEG) data. Unlike traditional brain connectivity networks, CNN-LSTM utilizes a self-attention mechanism to generate task-relevant connectivity graphs through end-to-end training. The method achieved a remarkable accuracy of 73.5%, outperforming conventional Convolutional Neural Networks (CNNs) and graph generation methods on a drowsy driving dataset. In recent years, the application of deep learning techniques to EEG signal classification has shown great promise, particularly in fields such as brain-computer interfaces and neurodiagnostic tools. This project explores the efficacy of two distinct neural network architectures, Interpretable Convolution Neural Network (InterpretableCNN) and Convolution Neural Network with Long Short-Term Memory (CNNLSTM), in classifying EEG signals. The dataset utilized for this study contains EEG samples from multiple subjects, each sample being a 3-second long segment recorded at a sampling frequency of 128 Hz. To ensure robustness and generalizability of the models, a leave-one-subject-out (LOSO) cross-validation strategy is employed. This involves training the model on data from all subjects except one, which is used for testing. This process is repeated for each subject, ensuring that the model's performance is evaluated comprehensively across different individuals.