Identifying Reproducibly Important EEG Markers of Schizophrenia with an Explainable Multi-Model Deep Learning...

Area of Science:

• Neuroscience
• Machine Learning
• Psychiatry

Background:

• Diagnosing schizophrenia (SZ) is complex due to varied symptoms.
• Explainable machine learning (ML) methods are used to find SZ biomarkers.
• The generalizability of biomarkers from ML studies is often limited by small model sample sizes.

Purpose of the Study:

• To develop novel methods for feature interaction-based explainability.
• To create approaches for summarizing multi-model explanations.
• To extract generalizable insights from ML models for SZ biomarker discovery.

Main Methods:

• Implemented a novel feature interaction-based explainability approach.
• Developed new methods for summarizing multi-model explanations.
• Analyzed electroencephalogram (EEG) spectral power data and model explanations (training and test sets).

Main Results:

• Identified significant effects of SZ on α, β, and θ frequency bands.
• Found SZ-related effects in the left hemisphere of the brain.
• Observed interhemispheric interaction differences associated with SZ across most cross-validation folds.

Conclusions:

• The developed methods enhance the reliability of biomarker identification in neuropsychiatric disorders.
• Findings provide insights into the neural underpinnings of schizophrenia.
• The study encourages the development of robust, explainable ML approaches for biomarker discovery.