A new diagnostic framework uses signal-processing mathematics to classify schizophrenia from EEG recordings and explain its own reasoning.
Researchers applied the Wavelet Scattering Transform to resting-state multichannel EEG data, extracting features that capture how brain signal amplitudes modulate across multiple timescales. They paired it with a strict leave-one-subject-out validation scheme — meaning the model was always tested on someone it had never seen — and added SHAP explainability to identify which features drove each prediction. A Random Forest classifier hit 90.48% accuracy with an AUC of 0.9339 and sensitivity of 95.56% under that rigorous evaluation. The electrode at position P3 emerged as the single most discriminative recording site.
The methodology critique embedded in the paper is as important as the results. Most prior EEG classification work uses epoch-level cross-validation, which can let fragments of the same subject's brain activity bleed between training and test sets — inflating reported accuracy without improving real-world utility. By enforcing subject-level splits, this work sets a higher bar that makes the 90% figure harder to dismiss. The dominance of gamma-band cross-frequency coupling features also gives clinicians a concrete electrophysiological signature to investigate, not just a black-box score.
Schizophrenia affects roughly 24 million people globally, yet diagnosis still depends almost entirely on clinical interviews. Objective EEG biomarkers have been a research goal for decades; the gap between lab accuracy and clinical adoption remains wide, and a single dataset result — however clean the methodology — is not a diagnostic tool yet.