The effect of acute stress on the interpretability and generalization of schizophrenia predictive machine learning models

Schizophrenia is a severe mental disorder, and early diagnosis is essential for improving patient outcomes. While ongoing research continues to advance understanding, the disorder’s complexity still limits accurate prediction. Combining electroencephalography (EEG) with machine learning (ML) has shown promise in aiding diagnosis, but overlapping mental health conditions such as stress can reduce the interpretability and reliability of ML models. This study utilized ML models using open EEG datasets to predict schizophrenia and investigate the impact of acute stress response during EEG recordings on model performance. Experiments utilized three open EEG datasets: acute stress, schizophrenia recorded at rest and schizophrenia recorded during tasks, both with healthy control subjects included. Four XGBoost-based classification models were developed: (1) acute stress, (2) schizophrenia at rest, (3) during tasks, and (4) a 3-class model that combined healthy controls with both schizophrenia groups. Explainable AI techniques were applied to further evaluate model performance against known schizophrenia brain region and frequency domain markers. A novel EEG artifact adjustment method for stress compensation was applied, and model performance re-evaluated. Results showed that acute stress response significantly affected EEG recordings and ML model accuracy, with compensation for acute stress improving model generalization. Findings underscore the need for rigorous health screening, artifact processing and stress response management during EEG recordings to ensure high-quality data for ML models. • We show that acute stress experienced during EEG recording may adversely affect machine learning model predictive accuracy, specifically models built to predict schizophrenia diagnosis. • We propose detecting acute stress within EEG data, and treating it as a physiological data artifact. • We further propose a novel method to remove these stress artifacts from EEG data, resulting in improved machine learning model generalization. • In favor of reproducible research and to advance the field, all programming code used in this study is made publicly available.