A three-dimensional STFT representation for EEG alcoholism classification using 3D convolutional networks

Abstract

Alcohol consumption can alter electrical activity in the brain, and these changes can be captured through electroencephalogram recordings. Automatically recognizing such patterns may help support early identification and reduce the reliance on subjective screening methods. In this work, a three-dimensional Short-Time Fourier Transform (3D STFT) representation is developed to encode temporal, spectral, and spatial characteristics contained in multichannel EEG data. Each electrode receives the STFT in order to produce a unique time-frequency map, which are then combined into a 3D tensor that preserves how the signal evolves across channels. Unlike 2D spectrogram, the 3D representation preserves spatial interactions across electrodes within the time-frequency domain. This multidimensional structure allows the model to interpret the EEG not as separate slices but as a unified volume. The 3D tensor is subsequently used as a 3D CNN's input for classifying EEG trials into alcoholic and non-alcoholic categories. Evaluation is performed using publicly available UCI Alcoholic EEG dataset. The proposed approach yields strong performance, producing accuracy, precision, and recall reaching 98.96%, 98.76%, and 99.17%, respectively, alongside an F1-score of 98.96 and AUC of 99.50%.. These results indicate that combining temporal, spectral, and spatial information within a single representation allows the network to extract deeper and more informative neural patterns compared to conventional 1D features or 2D spectrogram-based inputs. While STFT’s fixed window length limits its ability to represent rapid or highly irregular non-stationary changes, the overall results show that the 3D representation provides a comprehensive view of EEG dynamics and enables effective classification of alcoholic versus non-alcoholic subjects.