Researchers at IBM Research, working with university partners, have developed an artificial intelligence system that can predict the eventual onset of psychosis in at-risk patients by analyzing their speech patterns. The system, described in a study published in the journal World Psychiatry, achieved up to 83 percent accuracy in retrospective tests, according to the research team.
The work builds on a 2015 IBM study that first demonstrated the feasibility of using machine learning to distinguish speech differences between high-risk patients who later developed psychosis and those who did not. In that earlier research, the team quantified two clinical concepts—'poverty of speech' and 'flight of ideas'—using natural language processing (NLP) to measure syntactic complexity and semantic coherence.
For the new study, the researchers expanded their approach. Instead of asking patients to talk freely about themselves for an hour, as in the earlier work, participants were asked to retell a story they had just read. The AI was trained using insights from the 2015 study and then applied to the new, larger patient group.
According to the study, the system could have predicted which patients would eventually develop psychosis with 83 percent accuracy. When applied retrospectively to the original 2015 cohort, the model achieved 79 percent accuracy.
Toward Objective Psychiatric Assessment
Guillermo Cecchi, lead researcher and manager of the Computational Psychiatry and Neuroimaging groups at IBM Research, said the new system could help address the subjectivity inherent in traditional psychiatric evaluations. In a 2017 IBM Research post, Cecchi argued that AI and machine learning tools could make assessments more objective and improve diagnostic accuracy.
The potential clinical applications are significant. Cecchi told Futurism that the system could be used in clinics to quickly triage at-risk patients, allowing limited resources to be directed toward those most likely to experience a first episode of psychosis. For individuals without access to specialized professionals, the AI could potentially evaluate audio samples remotely.
The approach may extend beyond psychosis. Cecchi noted that similar methods could be applied to other conditions, such as depression. IBM Research is already exploring computational psychiatry for diagnosing and treating depression, Parkinson's disease, Alzheimer's disease, and chronic pain.
In a separate 2017 effort, Cecchi's team and researchers from the University of Alberta, working through the IBM Alberta Center for Advanced Studies, combined neuroimaging with AI to predict schizophrenia by analyzing brain scans.
While the findings are promising, the research is still in its early stages. The studies rely on retrospective analysis, and further validation in prospective clinical settings will be necessary before the tool can be widely adopted. Nonetheless, the work represents a step toward making neuropsychiatric assessment more accessible and data-driven.
