Early identification of schizophrenia remains one of psychiatry’s most difficult challenges. Without reliable tools, it can be hard to distinguish a single episode of psychosis from the early signs of a chronic, severe mental illness.
Now, a team of Columbia researchers is developing a deep learning model aimed at identifying patients with early psychosis who are at high risk of progressing to schizophrenia, using only information already found in standard electronic health records.
“If you’re a psychiatrist sitting with a patient experiencing their first episode of psychosis, you have to determine whether they’ll need long-term care,” says Steven A. Kushner, M.D., Ph.D., professor of psychiatry and co-director of the Stavros Niarchos Foundation Center for Precision Psychiatry & Mental Health at Columbia. “But right now, we don’t have a reliable way to accurately assess that future risk. An AI model could help clinicians make this critical determination earlier and with greater confidence.”
PET scan of the brain of a patient with schizophrenia having a hallucination.
What the AI Model Reveals
To assist clinicians in making earlier, more informed decisions, the researchers trained a deep learning model on Medicaid claims data, using a large dataset spanning years of longitudinal health records. What emerged as the most predictive features were not symptom reports but an increased frequency of healthcare services, including how often a person sought care, in which settings, and what services they received.
These patterns may reveal early clinical insights that aren’t always captured in standard psychiatric evaluations.
— Dr. Steven A. Kushner
“We had not expected that model performance would be so reliant upon patterns of healthcare use, including emergency room visits, hospitalizations, and outpatient appointments,” Dr. Kushner says. “These patterns may reveal early clinical insights that aren’t always captured in standard psychiatric evaluations.”
To ensure fairness within the model — one of AI’s known risks is replicating bias — Shalmali Joshi, Ph.D., assistant professor of biomedical informatics and member of the Data Science Institute at Columbia, has emphasized evaluating performance across different groups, such as patients with variable healthcare utilization patterns and other demographics.
In one analysis, the AI model showed better overall performance at predicting schizophrenia among women with early psychosis than for men, but its sensitivity was lower in identifying women at risk, resulting in more false negatives.
“AI models trained on real-world data can reflect the biases of the system they’re built from,” Dr. Joshi explains. “If we don’t examine how AI models perform across subgroups, we risk reinforcing existing biases rather than reducing them.”
We want to transform electronic health records from a source of administrative burden into a tool that actively supports better care.
— Dr. Shalmali Joshi
Clinical Validation and Future Steps
To test the model’s accuracy in clinical settings, the team is conducting chart reviews at NewYork-Presbyterian and Columbia. Each patient record is reviewed by three independent clinicians to evaluate the model’s predictions.
“This level of validation is essential before any clinical use,” says Dr. Joshi. “Ultimately, we want to transform electronic health records from a source of administrative burden into a tool that actively supports better care.”
The researchers also plan to integrate genetic data in future phases of the study, in collaboration with the New York Genome Center, to further refine risk predictions.
As development continues, Dr. Kushner emphasizes that the model is meant to be a tool for guidance, not a substitute for human expertise. By providing physicians with timely insights drawn from real-world data, he hopes it can support more precise, confident decision-making in complex cases.
Says Dr. Kushner, “We want to give psychiatrists better tools — not to replace clinical judgment, but to enhance it.”
