Schizophrenia Detection Using Convolutional Neural Networks on EEG Data

Abnormal corollary discharge has been implicated in schizophrenia and manifests as reduced suppression of auditory evoked responses during self-generated sounds. We investigate whether deep convolutional neural networks (CNNs) trained on EEG from a basic button–tone task can detect schizophrenia at the single-trial/subject level. We analyzed EEG from 81 participants (schizophrenia and healthy controls; combined across a prior publication and a larger replication cohort) collected during three conditions: (1) button press generating a tone, (2) passive tone, and (3) button press without tone. Preprocessing included re-referencing to averaged earlobes, 0.1 Hz high-pass filtering, canonical correlation analysis for muscle/high-frequency noise removal, ICA artifact rejection, epoching, baseline correction, and interpolation of outlier channels/trials. We trained a 2D-CNN optimized with Adam where inputs comprised from up to 64 electrodes. On a held-out validation set, the model achieved accuracy = 0.6265 and val loss = 0.6455. From the confusion matrix, we obtained recall (schizophrenia) = 0.578, specificity = 0.640, precision = 0.308, and F1 = 0.402.