Retracted: Deep Recurrent Architectures for EEG-Based Classification of Major Psychiatric Disorders: A Comparative Study of LSTM, BiLSTM, GRU, BiGRU, and Their Hybrids

Bipolar disorder (BD), major depressive disorder (MDD), and schizophrenia (SZ) are serious mental disorders that affect millions of people worldwide. Due to the overlapping clinical symptoms, early and accurate diagnosis of these disorders is difficult. This reduces patients’ quality of life, increases healthcare costs, and negatively impacts the treatment process. This study proposes an objective approach to support clinician decisions based on electroencephalography (EEG) signals and deep learning models. EEG signals recorded from 106 patients were divided into three groups: patients with SZ ( $n=43$ ), BD ( $n=30$ ), and MDD ( $n=33$ ). Feature extraction from the EEG signals was conducted in two different experiments. In the first experiment, preprocessed EEG data were directly applied to the input of deep learning models for automatic feature extraction. In the second experiment, short-time Fourier transform (STFT) was applied to the EEG signals, and the obtained features were used as input for classification. In order to evaluate the classification performance of the experiments, six different deep learning models, including unidirectional, bidirectional, and hybrid architectures, were used: gated recurrent unit (GRU), bidirectional GRU (BiGRU), long short-term memory (LSTM), bidirectional LSTM (BiLSTM), hybrid LSTM-GRU, and BiLSTM-BiGRU models. The experimental results showed that time-frequency domain features were more discriminatory than automatically extracted features. The BiLSTM-BiGRU hybrid model using the STFT features reached the highest classification performance with 94.55% accuracy, 0.944 recall, 0.972 specificity, 0.946 precision, and 0.945 F1-score. Our findings demonstrated the potential of deep learning models using time-frequency domain features as input in multi-class classification of mental disorders.