A deep learning model for clinical outcome prediction using longitudinal inpatient electronic health records.

Summary

TECO, a Transformer-based model trained on 2,579 hospitalized COVID-19 patients, consistently outperformed EDI, RF, and XGBoost for ICU mortality prediction and generalized to external ARDS-related MIMIC datasets. It also surfaced clinically interpretable, outcome-correlated features, suggesting utility as an early warning system across inpatient conditions.

Key Findings

• In development (COVID-19 cohort, n=2,579), TECO achieved AUC 0.89–0.97, surpassing EDI (0.86–0.95), RF (0.87–0.96), and XGBoost (0.88–0.96).
• In two external MIMIC test datasets, TECO yielded AUC 0.65–0.77, higher than RF (0.59–0.75) and XGBoost (0.59–0.74).
• The model identified clinically interpretable features correlated with mortality risk, supporting transparency and potential bedside adoption.

Clinical Implications

Hospitals could deploy TECO-like systems to flag high-risk ICU patients earlier, potentially guiding staffing, triage, and escalation of supportive therapies in ARDS and related conditions.

Limitations

• Proprietary EDI comparator unavailable in MIMIC, limiting head-to-head comparison
• Retrospective EHR design; prospective impact and generalizability require further validation

Future Directions

Prospective, multi-site deployment studies to assess clinical impact, calibration drift monitoring, fairness auditing, and integration with clinician workflows.