An Explainable and Robust Deep Learning Approach for Automated Electroencephalography-based Schizophrenia Diagnosis

Area of Science:

• Neuroscience
• Artificial Intelligence

Background:

• Schizophrenia (SZ) diagnosis relies on symptom variability, necessitating objective methods.
• Deep learning on electroencephalography (EEG) offers high temporal precision for automated SZ diagnosis.
• Model robustness and explainability are crucial for clinical application of AI in SZ detection.

Approach:

• Developed a novel channel dropout (CD) method integrated into a convolutional neural network (CNN) architecture (CNN-CD).
• Assessed model robustness against channel loss, a common issue in EEG recordings.
• Utilized explainability techniques to analyze spatial and spectral features learned by the models.

Key Points:

• The CNN-CD model demonstrated decreased sensitivity to channel loss compared to the baseline CNN.
• Both models achieved high diagnostic accuracies (CNN: 81.9%, CNN-CD: 80.9%) on testing data.
• Identified prioritization of parietal electrodes and the alpha-band, aligning with existing schizophrenia research.

Conclusions:

• The channel dropout approach enhances the robustness of explainable AI models for EEG-based schizophrenia diagnosis.
• Findings support the transition of AI tools from research to clinical decision support systems.
• Further development of robust and explainable AI is encouraged for improved psychiatric disorder diagnostics.