Poststroke Depression Detection Based on Multidistance Optimized Domain Adaptation Networks Using EEG Signals

Poststroke depression (PSD) is a common poststroke complication. Current studies have been conducted using electroencephalograph (EEG) signals for PSD detection and have achieved some advantages. However, there are many challenges in extracting common features of EEG signals across subjects in performing PSD detection. To overcome this problem, we propose a new multidistance optimized domain adaptation network (MODAN) model for PSD detection. MODAN extracts the common features of the EEG signal through a common feature extractor. Then, features of the hidden layer in the common feature extractor are fused with the output features to obtain multiscale features of the EEG signal to exploit both detailed and semantic features of the EEG signal. Finally, the source and target domain data are classified separately by a classifier to obtain the final detection results. In this process, multiple distance metrics are used by us to tune the model. The experimental results show that our proposed method achieves better detection results on all rhythms compared with the baseline systems. In addition, a new frequency band was identified that can be better used for the detection of PSD by our model. Furthermore, an ablation experiment was performed to validate the effectiveness of our proposed method. The method proposed by us is effective for PSD detection, and the new band identified by us is probably a new biomarker for PSD detection using EEG signals. This method provides a significant improvement in performance over the baseline systems.