Back to News
InnovationAI Understanding briefing

AI model predicts schizophrenia risk from Medicaid claims

A new study shows a deep‑learning model trained on Medicaid claims can estimate which patients with early psychosis will later be diagnosed with schizophrenia, achieving an AUROC of up to 0.79.

4 min readRead the linked source
Source-page capture accompanying AI model predicts schizophrenia risk from Medicaid claims
Source referenceSource recorded
Publisher
kevinmd.com
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

Key terms

Classifier
A model designed specifically for classification tasks.
Bias
A consistent pattern of error or unfairness in data or model behavior.

What happened

Columbia researchers trained a deep‑learning model on Medicaid claims from 158,704 individuals with a history of psychosis. The model identified patterns of health‑care utilization—diagnoses, prescriptions, visit frequency, and care settings—to predict later schizophrenia diagnosis. In the primary cohort, 8.5 % progressed to schizophrenia and the model reached an AUROC of 0.791; in a separate commercially insured cohort it achieved 0.763. The team is now reviewing patient charts at NewYork‑Presbyterian and Columbia to validate predictions before any clinical deployment.

The study, published in the journal *Schizophrenia* (doi:10.1038/s41537-026-00828-w), used de‑identified Medicaid claims from 158,704 individuals who had experienced psychosis. Researchers at Columbia University’s reAIM Lab built a deep‑learning that learned from patterns such as psychiatric diagnoses, medication prescriptions, frequency of visits, and care settings (ER, inpatient, outpatient).

Among the cohort, 8.5 % were later diagnosed with schizophrenia. The best performing model achieved an area under the receiver operating characteristic curve (AUROC) of 0.791 in the Medicaid sample and 0.763 in an external commercially insured cohort, indicating good discrimination between those who would and would not develop schizophrenia.

The authors emphasize that the model is intended to support, not replace, clinical judgment. It is not yet deployed in practice. The research team is currently reviewing patient charts at NewYork‑Presbyterian and Columbia to compare model predictions with actual clinical outcomes, a necessary step before any prospective use.

The study also reported lower sensitivity for women, suggesting that the model may miss a higher proportion of at‑risk female patients. The authors note that models trained on real‑world claims data inherit biases present in the health‑care system, such as under‑representation of individuals who never seek care.

Source details: kevinmd.com

Why it matters

Early identification of schizophrenia could transform care pathways for young people experiencing a first psychotic episode, a period when clinicians and families struggle to decide whether the episode is isolated or the start of a chronic illness. Reliable risk stratification would enable proactive treatment planning, targeted monitoring, and potentially reduce the duration of untreated psychosis, which is linked to poorer long‑term outcomes. However, the model relies on billing data, which captures health‑care encounters only after families have already sought help, and it shows lower sensitivity for women, highlighting risks and the need for complementary sources such as family observations.

Schizophrenia often emerges after a first psychotic episode, but clinicians lack objective tools to predict which patients will develop a chronic disorder. Early intervention can improve functional outcomes, reduce hospitalization rates, and lower long‑term health‑care costs.

A claims‑based AI model provides a scalable way to flag high‑risk patients across large health systems, potentially prompting earlier referral to specialized services, more intensive monitoring, or preventive pharmacologic strategies.

The reliance on billing data means the model only becomes aware of risk after a health‑care encounter, which may be too late to prevent early deterioration. Incorporating family‑reported changes—sleep disruption, social withdrawal, communication shifts—could improve timeliness and accuracy.

Gender in the model’s performance raises concerns about equitable care. If women at risk are systematically under‑detected, they could miss critical early interventions, exacerbating existing disparities in mental‑health outcomes.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

What to watch next

Future steps include prospective validation of the model in real‑time clinical settings, integration of structured family‑reported symptoms, and assessment of how risk alerts will be communicated to patients and caregivers. Monitoring for disparities—especially gender —and ensuring that any clinical decision‑support tool includes clear follow‑up plans will be critical before widespread adoption.

Prospective validation: Whether the model maintains its predictive performance when applied to live patient data in real‑time clinical workflows.

Integration of caregiver input: Trials that combine claims‑based risk scores with structured family observations could enhance early detection.

Implementation pathways: How health systems will present risk alerts to clinicians, ensure patient consent, and develop actionable care plans that involve families.

mitigation: Ongoing assessment of model performance across sex, race, socioeconomic status, and insurance type to prevent widening health inequities.

Related guides & quizzes

Found this useful?