New research from James Cook University shows that artificial intelligence can help clinicians distinguish between healthy people and those with schizophrenia using EEG brain-wave patterns, even when patients are under acute stress. The study, published in Biomedical Signal Processing and Control, details stress-adjusted machine learning algorithms that align with the latest diagnostic observations from medical science.
Schizophrenia affects roughly 1% of the global population and carries high mortality rates. Because many individuals show symptoms before psychosis begins, earlier and more accurate detection is critical. The multidisciplinary team-led by JCU Ph.D. candidate Gideon Vos and including engineers, data scientists, neuroscientists and psychology researchers-trained AI models on open-access EEG datasets to account for the distorting effects of stress on brain signals.
Why stress alters EEG signals in schizophrenia
Electroencephalography records electrical impulses across different brain regions during rest and cognitive tasks. Under stress, however, the brain's electrical activity shifts. In people with schizophrenia, that shift differs in measurable ways from the response seen in healthy individuals. "There are so many overlapping symptoms between acute stress response and people with schizophrenia," Vos said. "But there's something about the way that people with schizophrenia react to stress that is slightly different from what healthy people would experience under the same kind of situation."
Training AI models with real-world conditions
Many existing AI models for schizophrenia classification use EEG data collected in controlled lab settings-a process that itself induces stress because of the intrusive electrode setup and unfamiliar environment. Vos argued that ignoring this confound undermines model reliability. "EEG can be intrusive and uncomfortable, as it requires placing electrodes onto the patient's head, and you're in a lab rather than a home environment, so they're going to experience acute stress," he said. "It's important when building AI models that you understand the condition that they're building the model for."
The team's stress-adjusted algorithms produced physiologically consistent patterns that match established diagnostic knowledge. The models do not aim to replace clinical judgment; instead, they use explainable AI to surface the reasoning behind each classification.
Explainable AI opens the black box
Explainable AI requires models to show why they separate groups, not just that they can do so accurately. "If I train a model and the model says it can classify people perfectly into healthy and schizophrenic, most people would stop there and say, 'great, we've got a model that can predict schizophrenia,'" Vos said. "But we need the AI model to explain why it can, or why it can't, separate these two groups. We need to be able to take those explanations and have a scientific rationale for why they occur … by the actual medical professionals."
This approach opens a path to more accessible diagnosis. Explainable AI offers a pathway to more personalized and timely AI for Healthcare, particularly for people in remote and regional communities who may wait months for a specialist. "I can go to my GP tomorrow to get a diagnosis … but a lot of people are not that lucky. They might be hours away from even just a basic GP service," Vos said. "AI can be a tool on their smartphone that can see if something is potentially urgent, connecting them to a health care provider who can use the AI predictions and explanations in their own diagnostic process."
Why this matters for science and research professionals
For researchers and scientists working at the intersection of machine learning and clinical medicine, this study underscores a methodological imperative: models must be built with an understanding of the real-world conditions-like acute stress-that shape the data. The approach used in this research exemplifies the growing field of AI for Science & Research, where reproducibility and physiological grounding matter as much as raw accuracy. By requiring AI to explain its decisions in terms that clinicians can validate, the work provides a template for diagnostic tools that amplify, rather than override, professional expertise.
Your membership also unlocks:
