WideConvNet: A Novel Wide Convolutional Neural Network Applied to ERP-Based Classification of Alcoholics

This paper presents WideConvNet, a novel convolutional neural network (ConvNet or CNN) that outperforms three established end-to-end models – EEGNet, Shallow Convolutional Neural Network (ShallowConvNet), and Deep Convolutional Neural Network (DeepConvNet) – in classifying event-related potentials (ERPs) from electroencephalogram (EEG) signals. At the core of WideConvNet is an adapted inception module for 1D data, utilizing varying size 1D filters inspired by the wavelet scattering transform, which are learned during the training phase to capture temporal patterns across different scales within the EEG signals. The performance of WideConvNet was evaluated on multichannel EEG data from 20 male subjects with use disorder (AUD) and 20 healthy male controls, each providing 16 1.0-s ERPs. Using twice repeated 10-fold leave-subjects-out stratified cross-validation, WideConvNet largely outperformed both EEGNet and ShallowConvNet in terms of area under the curve of the receiver operating characteristic (AUC-ROC, p=0.00003,0.0001 ). Whereas, in terms of accuracy it largely outperformed EEGNet and moderately outperformed ShallowConvNet ( p=0.00029,0.00024 ). Despite comparable performances to DeepConvNet, WideConvNet required 28% fewer parameters and was more interpretable. The Shapley additive explanation (SHAP) values revealed parietal and occipital regions as the key contributors to both WideConvNet and DeepConvNet predictions. The findings suggest that WideConvNet, specifically on short-duration EEG signals, offers a balanced solution, combining robust classification performance with improved computational efficiency, making it suitable for clinical diagnosis and real-time brain-computer interface (BCI) applications.