Researchers used a mathematically grounded signal transform to find interpretable EEG biomarkers and to classify schizophrenia from resting brain recordings. They applied the multi-order Wavelet Scattering Transform (WST) to short epochs of 16-channel resting-state electroencephalography (EEG). Using a strict Leave-One-Subject-Out (LOSO) test, they report that a Random Forest classifier reached 90.48% accuracy (AUC = 0.9339; sensitivity = 95.56%), while also identifying concrete features linked to the diagnosis.

The study used a publicly available adolescent dataset of 84 subjects (45 with schizophrenia, 39 healthy controls). Recordings used a standard 16-electrode montage (including sites such as P3, Pz, Cz, F3) sampled at 128 Hz. Signals were cleaned with a 50 Hz notch filter and a 0.5–45 Hz bandpass filter, normalized per subject and channel, and cut into overlapping two-second epochs. The authors extracted three layers of WST coefficients: S0 (a slow baseline), S1 (energy at specific frequencies over time), and S2 (how fast those frequencies are amplitude-modulated by slower rhythms).

At a high level, the Wavelet Scattering Transform is a way to summarize how brain rhythms change and interact across time scales. In plain terms, S1 captures “how much” activity is in a frequency band, while S2 captures “how that activity is pulsed or driven” by slower rhythms. The authors tuned the transform to a 1.0-second averaging window and a frequency-resolution setting that favors fine spectral detail in the first layer and tight temporal localization in the second. They used subject-level analysis of variance (ANOVA) with Benjamini–Hochberg false discovery rate (FDR) correction to pick significant features, and SHapley Additive exPlanations (SHAP) to map model decisions back to individual scattering coefficients, electrodes, and frequency bands.