Schizophrenia Detection Using Convolutional Neural Networks on EEG Data

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning in Medicine

Background:

• Abnormal corollary discharge is linked to schizophrenia, evidenced by reduced auditory response suppression during self-generated sounds.
• Investigating novel biomarkers for schizophrenia detection is crucial for early diagnosis and intervention.

Purpose of the Study:

• To evaluate the efficacy of deep convolutional neural networks (CNNs) in detecting schizophrenia using electroencephalography (EEG) data.
• To assess the single-trial and single-subject detection capability of CNNs trained on a basic button-tone task.

Main Methods:

• EEG data were collected from 81 participants (schizophrenia patients and healthy controls) during three auditory conditions.
• Advanced preprocessing steps were applied, including filtering, artifact removal (ICA), and interpolation.
• A 2D-CNN model was trained using Adam optimization on EEG data from up to 64 electrodes.

Main Results:

• The CNN model achieved an accuracy of 0.6265 on a held-out validation set.
• Key performance metrics included recall (schizophrenia) = 0.578, specificity = 0.640, precision = 0.308, and F1 score = 0.402.
• The model demonstrated moderate performance in distinguishing schizophrenia from healthy controls at the single-trial/subject level.

Conclusions:

• Deep convolutional neural networks show promise as a tool for detecting schizophrenia from EEG data.
• The findings suggest that EEG patterns during a simple auditory task, analyzed by CNNs, may serve as potential biomarkers for schizophrenia.
• Further research with larger datasets and refined models is warranted to improve diagnostic accuracy.