Schizophrenia risk prediction following antecedent psychosis using deep learning

Schizophrenia is a debilitating psychiatric disorder in which earlier clinical recognition would enhance monitoring and treatment planning. We developed a deep learning model using administrative claims data to predict diagnostic transition to schizophrenia among individuals with a history of psychosis. We included 158,704 individuals from a Medicaid claims database (2008–2022) with a history of psychosis, of whom 13,437 (8.5%) were subsequently diagnosed with schizophrenia. Individuals with at least 3 years of continuous observation and at least one non-schizophrenia psychosis diagnosis prior to the censor date were included in the cohort. We used a validated phenotype to predict the individual-level likelihood of future incipient schizophrenia by leveraging temporal patterns of clinical features. Diagnostic transition was predicted using structured features, including conditions, medications, laboratory events, procedures, and visit timing. An encoder-only transformer architecture trained to learn the temporal signal of a patient’s trajectory performed the best on internal validation data compared to other architectures. External validity of this model was assessed using an independent, commercially insured cohort. The transformer model exhibited an AUROC of 0.791 [95% CI: 0.788, 0.794] and transported well to the external dataset (AUROC: 0.763 [0.751, 0.772]). Psychiatric care and healthcare utilization features were assessed to be the most predictive. Overall, we find promising performance for schizophrenia risk stratification among individuals with a history of psychosis. Features related to psychiatric care, particularly prescriptions and diagnoses, and healthcare utilization in outpatient settings were the most predictive of a future schizophrenia diagnosis.

Authors

Institutions

Publication Details

Journal
Schizophrenia
Published
2026-11-19
DOI
https://doi.org/10.1038/s41537-026-00828-w
Primary Topic
Machine Learning in Healthcare
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Schizophrenia risk prediction following antecedent psychosis using deep learning

Andrey Rzhetsky, Todd Lencz, Steven A. Kushner, Sander Markx et al.
Schizophrenia
Machine Learning in Healthcare
article

Schizophrenia risk prediction following antecedent psychosis using deep learning

Andrey Rzhetsky, Todd Lencz, Steven A. Kushner, Sander Markx, Michael C. Zody, Molly T. Finnerty, Noémie Elhadad, Aparajita Kashyap, Vaidehi Jobanputra, Paul S. Appelbaum, Shalmali D. Joshi, Matthew S. Lebowitz, Thomas Lehner, COPPER Working Group, Caline Karam
article en

Abstract

Schizophrenia is a debilitating psychiatric disorder in which earlier clinical recognition would enhance monitoring and treatment planning. We developed a deep learning model using administrative claims data to predict diagnostic transition to schizophrenia among individuals with a history of psychosis. We included 158,704 individuals from a Medicaid claims database (2008–2022) with a history of psychosis, of whom 13,437 (8.5%) were subsequently diagnosed with schizophrenia. Individuals with at least 3 years of continuous observation and at least one non-schizophrenia psychosis diagnosis prior to the censor date were included in the cohort. We used a validated phenotype to predict the individual-level likelihood of future incipient schizophrenia by leveraging temporal patterns of clinical features. Diagnostic transition was predicted using structured features, including conditions, medications, laboratory events, procedures, and visit timing. An encoder-only transformer architecture trained to learn the temporal signal of a patient’s trajectory performed the best on internal validation data compared to other architectures. External validity of this model was assessed using an independent, commercially insured cohort. The transformer model exhibited an AUROC of 0.791 [95% CI: 0.788, 0.794] and transported well to the external dataset (AUROC: 0.763 [0.751, 0.772]). Psychiatric care and healthcare utilization features were assessed to be the most predictive. Overall, we find promising performance for schizophrenia risk stratification among individuals with a history of psychosis. Features related to psychiatric care, particularly prescriptions and diagnoses, and healthcare utilization in outpatient settings were the most predictive of a future schizophrenia diagnosis.

Schizophrenia
Northwell Health (US), University of Chicago (US), New York Genome Center (US), Igenbio (United States) (US), New York University (US), Columbia University (US)
Machine Learning in Healthcare
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon