Artificial intelligence EEG biomarkers for schizophrenia diagnosis: Systematic review with PROBAST+AI assessment.

In 30 seconds

This systematic review assessed the diagnostic performance and methodological quality of AI models using resting-state EEG to identify schizophrenia or first-episode psychosis. It included 43 studies with 3,420 patients and 3,164 healthy controls. Mean reported accuracy was 93.38%, but performance differed by validation approach: 88.22% with subject-wise versus 98.14% with segment-wise validation. External validation was reported in only 4.7% of studies, and 76.7% were judged at high risk of bias, raising concern that data leakage inflated accuracy.

Key findings

• Mean diagnostic accuracy across studies was 93.38%; deep learning accuracy was 96.83% versus 87.68% for classical machine learning.
• Accuracy was 88.22% with subject-wise validation and 98.14% with segment-wise validation, a discrepancy the review linked to data leakage.
• Only 4.7% of studies used external validation, and 76.7% were assessed as having high risk of bias.

Why it matters

High reported accuracy alone may not indicate reliable diagnostic performance when validation methods permit data leakage or lack independent testing. The review highlights that methodological quality is central to interpreting AI-EEG results and to judging whether they can generalize beyond the datasets used to develop them.

What to keep in mind

The review reports substantial methodological heterogeneity among included studies, limiting confidence in aggregate performance estimates.