Deep Learning for Individual-Level Classification of Schizophrenia Versus Healthy Controls from Trial-Level Auditory Oddball ERP Waveforms

Machine learning approaches may support individual-level classification in psychiatry, but many EEG-based schizophrenia studies have relied on small samples or conventional summary features. We evaluated whether trial-level auditory oddball event-related potential (ERP) waveforms could support schizophrenia versus healthy-control classification using deep learning. The study included 258 patients with schizophrenia and 142 healthy controls. EEG recordings from an auditory duration oddball paradigm were segmented into -100 to 500 ms epochs, and trial-level mismatch waveforms were generated by subtracting each participant's mean standard response from accepted deviant trials. Models were trained using a fixed participant-level training, validation, and test split, with demographic residualization fit only in the training set. Five deep learning architectures were trained on full residualized ERP waveforms and compared with classical machine learning models trained on 18 conventional ERP summary features. Deep learning models achieved higher test set discrimination than classical feature-based models, with AUROC values ranging from 0.797 to 0.857 versus 0.705 to 0.720. Benchmark analyses suggested that performance depended on the combination of waveform-level input and deep learning architecture. These findings support trial-level auditory oddball ERP waveforms as promising classification inputs and candidate electrophysiological biomarkers of schizophrenia-related neural information processing.