Multiclass EEG-Based Classification of Psychiatric Disorders Using Interpretable Machine Learning on Spectral and Connectivity Features

Abstract

Psychiatric diagnosis relies heavily on subjective clinical evaluation, limiting objective differentiation across co-occurring disorders. This study presents an interpretable multiclass electroencephalography (EEG)-based framework for —Addictive Disorder, Anxiety Disorder, Mood Disorder, Obsessive-Compulsive Disorder (OCD), Schizophrenia, Trauma and Stress-Related Disorder—and Healthy Controls. A dataset of 945 subjects with 1,140 EEG-derived spectral and coherence features was analysed, with variance-based selection reducing the feature space to 300 descriptors. Classification was performed using Extreme Learning Machine (ELM), Naïve Bayes (NB), and Support Vector Machine (SVM) within a one-vs-all architecture. NB demonstrated comparatively limited discriminative capability (F1: 35.33%–72.76%). ELM achieved stable performance with F1-scores ranging from 66.03% to 89.31% and consistently higher testing accuracies across all classes, including 95.12% for OCD and 90.34% for Healthy Controls. SVM yielded the highest precision–recall balance across most classes, achieving F1-scores of 94.00% for OCD, 84.08% for Anxiety Disorder, and 79.84% for Schizophrenia. Mood Disorder remained the most challenging class (maximum F1: 66.03%). Bandwise analysis identified theta-band features as the most discriminative within the present dataset , achieving 91.57% standalone accuracy. These results suggest that EEG-derived spectral and coherence features, when combined with lightweight machine learning classifiers, can support multiclass psychiatric classification with variable class-wise performance. However, the lower performance observed for Mood Disorder and the reliance on a single public dataset indicate that further external validation, statistical comparison, and multimodal feature integration are required before clinical translation.