ISSN: 2161-1025
Maryam Saeedi, Pooya Mohammadi Kazaj, Abdolkarim Saeedi, Arash Maghsoudi, Alireza Vafaei Sadr
Background: Schizophrenia is a chronic mental illness in which a person’s perception of reality is distorted. Early diagnosis can help to manage symptoms and increase long-term treatment. The Electroencephalogram (EEG) is now used to diagnose specific mental disorders. Methods: In this paper, we developed an artificial intelligence methodology built on deep convolutional neural networks and transformer layers to detect schizophrenia from EEG signals directly, recordings include 14 paranoid schizophrenia patients (7 females) with ages ranging from 27 to 32 and 14 normal subjects (7 females) with ages ranging from 26 to 32. In the first phase, we used the Gramian Angular Field (GAF), including two methods: The Gramian Angular Summation Field (GASF) and the Gramian Angular Difference Field (GADF) to represent the EEG signals as various types of images. Then, well-known architectures, namely transformer and CNN-LSTM, are applied in addition to two new custom architectures. These models utilize two-dimensional Fast Fourier Transform layers (CNN-FFT) and wavelet layers (CNN-Wavelet) to extract useful information from the data. These layers allow automated feature extraction from EEG representation in the time and frequency domains. Ultimately these models were evaluated using common metrics such as accuracy, sensitivity, specificity and f1-score. Results: Transformer and CNN-LSTM models derive the most effective features from signals based on the findings. Transformer obtained the highest accuracy of 98.5 percent. The CNN-LSTM, which has a 95.7 percent accuracy rate, also performs admirably. Conclusion: This experiment outperformed other previous studies. Consequently, the strategy can aid medical practitioners in the automated detection and early treatment of schizophrenia.
Published Date: 2025-02-12;