Neuropsychiatric disorders constitute a substantial global health burden, affecting approximately 970 million individuals worldwide, with a exacerbation during the COVID-19 pandemic. This investigation presents a comparative analysis of three state-of-the-art artificial intelligence architectures–Multilayer Perceptron (ANN), TabNet, and FT-Transformer–for the automated classification of neuropsychiatric disorders using electroencephalographic (EEG) signals. The study employed a clinically validated dataset comprising 945 participants stratified across seven diagnostic categories: mood disorders, addictive disorders, trauma and stress-related disorders, schizophrenia, anxiety disorders, healthy controls, and obsessive-compulsive disorder. The neurophysiological data, already pre-processed in a previous study, were used in conjunction with the Synthetic Minority Over-sampling Technique (SMOTE) due to class imbalance. The dataset has the features: power spectral density (PSD) for six frequency bands, functional connectivity (FC), age, sex, and intelligence quotient (IQ). Experimental results demonstrated TabNet’s superior generalization capability across multiple psychiatric disorders, achieving notable performance in mood disorders (93.50% accuracy) and addictive disorders (96.00% accuracy). Our proposed ANN architecture exhibited exceptional discriminative power for schizophrenia classification (94.38% AUC) and obsessive-compulsive disorder detection (98.89% AUC). The FT-Transformer model yielded competitive performance, particularly in anxiety disorder identification. These findings underscore the critical importance of disorder-specific model selection and feature engineering, advancing the development of robust computational frameworks for neuropsychiatric diagnostic support systems.