Unified hierarchical learning framework for alcoholic EEG classification integrating an attention-based deep belief network with transformer-based classification models

This is the authors' abstract. We don't add key points for this paper.

Alcoholism causes profound structural and functional changes in brain activity that can be effectively observed via Electroencephalography (EEG). The early detection and prevention of serious neurological damage require reliable detection of alcoholic EEG patterns. However, EEG signals are extremely nonlinear, non-stationary and high-dimensional and proper classification is a difficult task. In this work, a comprehensive hierarchical learning framework is developed for alcoholic EEG classification. Multiple state-of-the-art feature extraction methods are used, including ABC clustering, WVD, FWHT, and Hybrid WOA-PSO to identify different structural, spectral, and time-frequency features of multichannel EEG signals. There are two experimental settings investigated. Transformer-based models use the extracted features directly to classify them in Experiment I. Experiment II proposes an A-DBN used to evaluate and filter the most discriminative features and then perform classification with transformer-based models. The experimental results indicate that WVD-based features are better in discriminative power compared to the other extraction methods. In Experiment I, the WVD + Neuromorphic Transformer model achieved the highest classification accuracy of 97.13%. Following attention-based refinement of the features, the WVD + A-DBN + Neuromorphic Transformer setup in Experiment II demonstrated a much higher accuracy of 99.48%, as well as increased stability and generalization. The findings indicate that the use of a combination of sophisticated time-frequency feature extraction and attention-based hierarchical feature selection and transformer-based learning is significantly more effective in enhancing alcoholic EEG classification. The proposed framework provides an effective and scalable solution to the automated diagnosis of alcohol-related brain abnormalities.