Comparison of Explainable Deep Learning Methods for Schizophrenia Detection

Schizophrenia is a severe mental disorder for which early, objective diagnosis is crucial, yet conventional clinical assessment can be subjective and inconsistent. This study presents an auto-mated EEG-based framework for schizophrenia detection that combines deep learning with explainable AI. EEG recordings acquired from 30 participants (15 patients, 15 healthy controls) across 16 chan-nels were preprocessed using the Douglas–Peucker algorithm, compressing data size by 40.15% while preserving key signal structure, and then transformed into time–frequency spectrograms via Short-Time Fourier Transform. Multiple transfer-learned architectures spanning residual CNNs, modern convolutional designs, and transformer-based models were systematically benchmarked under subject-level 5-fold cross-validation to prevent data leakage from the one-to-many subject-to-spectrogram mapping. Under subject-level partitioning, ConvNeXt achieved 88.56% ± ~9.46 % accuracy with 92.56% ± ~10.30 % AUC-ROC, with all five non-hybrid architectures achieving mean AUC-ROC above 91% on this preliminary dataset. Grad-CAM explanations revealed attention patterns broadly consistent with known neurophysiological markers, exposing architectural differences between convolutional and transformer models. These proof-of-concept findings suggest that time–frequency EEG modeling with interpretable deep learning is a promising research direction. However, multi-site validation on larger and more diverse cohorts is required before its generalizability and potential clinical applicability can be established.