EXPLAINABLE EEG-BASED SCHIZOPHRENIA DETECTION USING ANOVA FEATURE SELECTION AND OPTIMIZED EXTRA TREES WITH STRATIFIED CROSS-VALIDATION

Abstract

Schizophrenia is a chronic psychiatric disorder where the perception, cognition and behaviors are disturbed, which requires a proper and timely diagnosis to give an effective clinical intervention. EEG has been a recently viable non invasive technique for identifying neurophysiological markers related to schizophrenia. EEG-derived features, however, are usually very high dimensional and include redundant features and reduce classification performance. We propose a novel explainable machine learning approach for automated schizophrenia detection using EEG features. A publicly available EEG Psychiatric Disorders dataset was used to extract 1,144 features from both the demographic information and the EEG. To minimize feature redundancy, Analysis of Variance (ANOVA) based SelectKBest feature-selection was used to select 50 most discriminative-features. The hyperparameters of the Extra Trees classifier were then optimized by grid search and the selected features were classified using this optimized Extra Trees classifier. The model performance was assessed by an independent hold-out test set and Stratified-5-Fold Cross-Validation. The optimized model attained 95.35%, 95.29%, 95.83%, and 96.27% for hold-out test accuracy, balanced accuracy, F1-score and ROC-AUC, respectively. Moreover, the Stratified-5-Fold Cross-Validation provided a mean accuracy of 85.37 ± 5.01% and a mean ROC-AUC of 93.14 ± 2.51%, showing good generalization across the different data folds. Additionally, SHapley Additive exPlanations (SHAP) analysis revealed the most important demographic and EEG biomarkers for classification, improving the model interpretability. The results show that integrating ANOVA-based feature-selection with an optimized Extra Trees classifier, significantly improves classification performance while providing clinically interpretable predictions.