Multi domain feature fusion model for EEG based biomedical signal analysis in depression detection with a reduced electrode configuration

Abstract Depression is a neurological disorder, and biomedical signal analysis can support its diagnosis. Electroencephalography (EEG) is one of the biological signals that is employed to capture brain neural activity via multiple electrodes. However, the use of many electrodes increases both measurement and system costs and increases patient discomfort during EEG recording. This research introduces a model that employs a reduced number of electrodes to detect depression, incorporating the Fisher score algorithm to perform electrode reduction. It selects a subset of electrodes by focusing on those with Fisher scores that exceed both the mean and the mean+standard deviation (mean+SD) of Fisher scores across all electrodes, resulting in a reduced electrodes set. The proposed model uses time and frequency domain features. Spatial and temporal features are extracted using the self-attention mechanism in time domain. The EEG data undergoes a transformation into a frequency domain employing the fast Fourier transform, enabling the extraction of frequency-specific features. A novel approach is presented here, employing the fusion of time and frequency domain features, resulting in a comprehensive multi-domain feature set. The proposed TimeFreq-AttnNet model significantly reduces the count of electrodes needed, while preserving its effectiveness in detecting depression. The proposed model achieves 99.08% and 98.46% accuracy on the Hospital University Sains Malaysia (HUSM) and PREDICT datasets, respectively, using the mean threshold. With the mean+SD threshold, it attains 93.04% accuracy on HUSM and 94.43% on PREDICT.