Evaluation of a Generative Medical Event Foundation Model for Predicting Post-Discharge Trajectories in Emergency Department Abdominal Pain.

OBJECTIVE

Emergency department (ED) risk models predict whether a patient will revisit within a fixed window, but not what the visit will entail. We evaluated whether a medical event foundation model predicts the timing, sequence, and severity of downstream care after ED discharge for abdominal pain.

MATERIALS AND

METHODS

Retrospective cohort study in the de-identified Epic Cosmos network. From 150,030 eligible adults, we analyzed a random sample of 3,000, generating 50 simulated 30-day trajectories per patient with Curiosity, a generative medical event foundation model, and compared these with gradient boosted tree baselines. Endpoints were discrimination for three outcome classes (any revisit, discharge-revisit, admit-revisit) at three horizons, 36-class trajectory accuracy, transition-level calibration, and subgroup performance.

RESULTS

Curiosity's 30-day admit-revisit AUROC was 0.83 (95% CI 0.79-0.87) versus 0.76 (0.71-0.81) for XGBoost (P=.002). Differences were significant at five of nine outcome-horizon pairs but not for discharge-revisit at any horizon. Its most likely trajectory (of 36) matched the observed in 45.9% versus 43.0% (P<.001); median absolute calibration error across 45 transitions was 1.30 percentage points. Discrimination and calibration were poorer for Hispanic or Latino, self-pay, and other-race patients.

DISCUSSION: Without task-specific training, one pretrained foundation model matched or exceeded a tuned supervised baseline, with significant advantages on the primary endpoint and trajectory prediction but none for discharge-revisit. Subgroup gaps must be resolved before clinical use.

CONCLUSION

A medical event foundation model can represent post-ED-discharge revisit risk as a calibrated trajectory rather than a single probability; prospective evaluation and remediation of subgroup performance gaps remain necessary.