A comparative evaluation of EEG-based deep learning models for schizophrenia detection with cross-dataset validation and explainable AI

OBJECTIVES: Schizophrenia is a neuropsychiatric disorder that affects emotional, behavioral, and brain functions that can be tracked using electroencephalography (EEG). This research conducts a comparative evaluation of deep learning models utilizing EEG time-frequency and spectral analysis methods to automate schizophrenia detection. METHODS: Two compatible EEG datasets were merged, yielding a total of 934 EEG samples from 237 subjects (121 schizophrenia patients and 116 controls). Independent Component Analysis (ICA) was applied for signal decomposition. By deriving time-frequency representations using Continuous Wavelet Transform (CWT) and Fast Fourier Transform (FFT), scalogram and spectral inputs for deep learning models were obtained. Six architectures, including CNN variants, CNN-FFT, CNN-ELM, CNN-LSTM, ResNet Transfer, and a Transformer-based model, were evaluated with data augmentation and class balancing to improve robustness. RESULTS: While variations in numerical performance were observed across models, statistical analysis indicated that these differences were not significant. DISCUSSION: The study presents results that underscore the benefits of combining time-frequency analysis with deep learning for EEG-based schizophrenia diagnosis, especially via spectral feature extraction in CNN architectures. Furthermore, it provides insights consistent with known neurophysiological patterns in schizophrenia, emphasizing the significance of model interpretability for clinical translation. Future research will focus on the integration of multimodal neuroimaging and the enhancement of explainability frameworks to augment diagnostic reliability.