Can a machine learning model tell a schizophrenic patient from a healthy person by their brain’s electrical response to a visual stimulus? This project explores that question on real clinical data: EEG recordings of evoked potentials from the P300 visual paradigm, in a dataset of 54 patients and 54 healthy controls matched by age and sex.
The raw brain signals are too noisy to classify directly, so the first step pulls structure out of them: a discrete wavelet transform (Daubechies wavelets) breaks each signal into frequency components across several scales, yielding 180 features per subject. A Support Vector Machine trains on those features, tested with cross-validation.
The model reaches about 63% accuracy, a modest but real signal on a hard problem, and a full paper covers the method, results, and limits.