Summary
A medRxiv preprint describes MINT, a foundation model trained and validated on 766,733 pediatric emergency visits. The authors report advantages over task-specific models in 21 of 25 comparisons and stronger performance than physicians on one respiratory-support forecasting task.
A medRxiv preprint describes MINT, a foundation model designed to learn from electronic health records (EHRs) at minute-scale resolution and forecast events in pediatric emergency care. The researchers report that it outperformed task-specific models in 21 of 25 comparisons and performed better than physicians on a task forecasting escalations in respiratory support.
The model was pretrained and validated using 766,733 pediatric emergency-department visits across five health systems and 10 hospitals, spanning 16 years of data. The findings are reported in a preprint, which has not completed peer review.
Why the time scale matters
EHR models are often trained on information grouped over hours, days or years. The authors argue that these longer time horizons do not provide the resolution needed to inform decisions at the bedside, where a patient's condition may change over a much shorter period. MINT’s minute-scale approach is intended to capture those finer-grained patterns in emergency-care records.
The study reports stronger performance and better generalisation to external health systems than task-specific models across 21 of 25 comparisons. It also reports that MINT outperformed physicians when forecasting escalation of respiratory support. That finding applies to this particular forecasting task; it is not a comparison of overall clinical performance.
The authors also describe capabilities including department-scale monitoring, dynamic risk explanations, forecasts of individual physiological responses and hypothesis generation. These point to possible uses beyond predicting outcomes for a single patient, although the preprint’s abstract does not provide the details needed to assess each capability separately.
Evidence and data access
The reported results come from model development and validation using historical emergency-department visits, not a prospective clinical trial. The abstract gives the number of superior comparisons but does not report effect sizes or task-by-task measures, which limits interpretation of the magnitude of the performance differences.
The University of California does not permit public release of the de-identified data used in the study. The authors say code will be made publicly available after peer-reviewed publication; distribution of model weights will follow institutional and regulatory policy. These access conditions matter for independent examination of the model and its reported results.