An EHR-only model was enough to predict clozapine initiation within 1 year among adults with schizophrenia or schizoaffective disorder in routine psychiatric care, with an AUROC of 0.81 in a held-out test set. In this Denmark-based study of 229,761 hospital visits across 5,806 patients, the best XGBoost model achieved 42% sensitivity and 23% positive predictive value at a 7.5% predicted positive rate.

Why It Matters To Your Practice

• Clozapine is the only proven treatment for treatment-resistant schizophrenia, but initiation is often delayed by years.

• A dynamic model using routinely collected structured EHR data and clinical notes could help flag patients who may be approaching clozapine eligibility sooner.

• For clinicians interested in AI, this is a practical example of prediction built from data already generated in care — no extra testing required.

Clinical Implications

• The model made predictions at each psychiatric hospital visit, estimating whether a first clozapine prescription would occur within the next 365 days.

• Inputs included 179 structured predictors plus 750 note-derived features, spanning diagnoses, medications, and coercive measures.

• Used as clinical decision support, such a tool could prompt earlier review of treatment resistance, adherence, and clozapine candidacy.

• Performance was not strong enough for autonomous decision-making alone, but it may be useful for case-finding and prioritizing chart review.

Insights

• The study included all adults in contact with the Psychiatric Services of the Central Denmark Region from 2013 to 2024 who had schizophrenia or schizoaffective disorder diagnoses.

• Models were trained on 194,234 visits from 4,928 patients and tested on 35,527 visits from 878 patients.

• The best-performing approach was XGBoost, evaluated against logistic regression using a held-out 15% test set.

• The signal came solely from routine EHR data, suggesting meaningful practice patterns and illness trajectories can be captured without bespoke data collection.

The Bottom Line

• AI built from everyday EHR data can identify patients likely to start clozapine within a year with good discrimination.

• For practice, the near-term opportunity is earlier recognition of possible treatment-resistant schizophrenia — not replacing clinician judgment, but supporting it.