Schizophrenia scrambles how brain networks coordinate, but there's still no solid EEG test for it, partly because common pipelines lean on static power-spectrum features that miss the dynamic stuff, and partly because sloppy validation lets the same person's data leak into both training and testing, inflating scores. This paper tackles both. It uses a wavelet scattering transform to capture amplitude modulation and cross-frequency coupling, and enforces strict leave-one-subject-out testing so no subject appears on both sides.

The results point to something specific. Second-order scattering features, the ones encoding cross-frequency coupling, dominated the useful biomarkers, with gamma-band activity most prominent and electrode P3 the single most informative site. A random forest reached about 90% accuracy with high sensitivity. Because they used SHAP for explanation, the markers are interpretable, not just predictive.

This rests on a cut-off abstract, so the paper carries the full biomarker list and validation.