research to bring this vision closer to fruition. They are both faculty at the Columbia University Vagelos College of Physicians and Surgeons.

“We train the AI models using a massive dataset that includes the full breadth and depth of information in each person’s longitudinal health record,” says Kushner. “What’s striking is that the most predictive factors identified by the model are patterns of healthcare use; specifically, the frequency and types of services a person receives, rather than solely medical symptoms.”

Among those factors, frequent emergency room visits, hospitalizations, and outpatient appointments – whether related to mental health or other medical issues – consistently stand out. Traditionally, these aspects of care utilization have not been part of the formal diagnostic criteria for schizophrenia. But the AI model’s ability to systematically profile this information offers a more refined view, potentially enabling earlier and more accurate diagnoses.

Ensuring Generalizability in AI Models

A critical challenge in this research is ensuring that AI tools are both accurate and generalizable across patient populations. Joshi, who studies how machine learning models can be influenced by uneven or incomplete data, explains, “If an algorithm is trained on data that disproportionately represents certain subpopulations, the algorithm may pick up and amplify existing associations dominant in that subpopulation that may not exist in other groups – for example, patients who use healthcare less than the average utilization among patients with psychosis.”

At Columbia’s Data Science Day 2024, Joshi presented findings showing that an AI model designed to predict schizophrenia in a Medicaid cohort representing a few U.S. states, is more accurate for women than for men among patients with early psychosis. The model, however, was less sensitive in identifying women at risk compared to men and has a higher specificity for women than men. More generally, the model exhibits a wider sensitivity-specificity tradeoff for women than for men. This underscores the need for meticulous evaluation and adjustment of AI models to ensure they serve all populations effectively, Joshi says.​

This research reflects the goals of Columbia AI, a university-wide initiative housed at the Data Science Institute at Columbia University to advance the responsible development and deployment of artificial intelligence across disciplines. Faculty affiliated with the Data Science Institute spans all disciplines at Columbia, and collaborations like this are central to the Institute’s mission of “Data for Good.” Joshi’s affiliation with the Institute reflects its interdisciplinary approach to AI in health, and its commitment to building models that work for diverse patient populations. 

The Power of Early Prediction

Schizophrenia can present with a range of symptoms, including hallucinations, delusions, and disorganized thinking. But in its early stages, it often resembles other mental health conditions, making diagnosis especially challenging. “The most common first diagnosis for people who eventually develop schizophrenia is depression,” explains Kushner. “That doesn’t mean clinicians are making mistakes; it’s simply that early schizophrenia symptoms can look very similar to depression.”

Unlike many areas of medicine, psychiatry lacks objective tests – no bloodwork, brain scans, or genetic screens can confirm a diagnosis. Clinicians instead rely on patient history, reported symptoms, and clinical observation, which can make it difficult to detect when someone is on the path toward a serious mental illness.

That uncertainty can delay the start of appropriate care. And for serious mental illnesses (SMIs), which affect approximately 1 in 17 adults in the U.S., such delays can have lasting consequences. Research consistently shows that the earlier schizophrenia is identified and treated, the better the long-term outcomes.

“If you’re a psychiatrist sitting with a patient who has experienced psychosis, you have to make a crucial decision – do they need long-term intensive care or is this a one-time episode?” says Kushner. “Right now, about 8% of people in the Medicaid dataset with psychosis eventually develop schizophrenia, but we don’t yet have a clear way of identifying who is at risk. It would be very beneficial for optimal treatment planning to know who those people were earlier in the process.”

Validating Through Chart Reviews

Joshi and her team are currently training new AI models on a cohort of patients receiving care at Columbia Irving Medical Center. Further, Joshi and team will conduct chart reviews to validate the models’ accuracy and reliability in real-world clinical settings by ensuring that each chart is reviewed by three practicing clinicians.

“Advancing the development of AI models to address multiple clinically meaningful tasks at a time, including scaling the chart review and validation process, has the potential to decrease the time providers spend on manual auditing,”Joshi says. “It can also identify key information, generate summaries, and enhance decision-making through insights drawn from large datasets, thereby turning electronic health records from a source of administrative burden into a valuable tool for improving patient care.”​

Advancing Precision Psychiatry

Central to these initiatives is Columbia’s Center for Precision Psychiatry & Mental Health, established through a $75 million grant from the Stavros Niarchos Foundation. The Center is co-directed by Kushner, Sander Markx, and Joseph Gogos, bringing together leaders in psychiatry, neuroscience, and biophysics from across Columbia and the Zuckerman Mind Brain Behavior Institute.

The Columbia SNF Center is focused on driving scientific innovation and bringing the tools of precision medicine into everyday psychiatric care. Rather than relying on the traditional trial-and-error approach, the Center focuses on understanding the genetic, biological, and environmental factors that contribute to each person’s mental health, ultimately allowing for more targeted and effective treatments.

Looking ahead, Kushner and Joshi plan to expand their work. Once the retrospective analyses are complete, they will begin the next phase of the COPPER grant. In partnership with the New York Genome Center, they will collect genetic data from a large cohort of patients to determine whether the inclusion of genetic information further strengthens their models.  They will also explore the ethical, legal, and societal questions that come with integrating AI and genetic data into mental health care.

The ultimate goal is to develop an AI-driven tool that can assist any psychiatrist in making a more accurate diagnosis from the very first patient visit. “The aim isn’t to replace clinicians, but to give them a powerful way to make better-informed decisions,” says Kushner.  “Interdisciplinary approaches like this are clearing the path to more timely, more personalized, and ultimately better mental health care for all.”