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

Summary

The TECO Transformer model trained on 2,579 COVID-19 inpatients achieved AUC 0.89–0.97 for ICU mortality prediction and outperformed EDI, RF, and XGBoost. External validation in ARDS (n=2,799) and sepsis (n=6,622) cohorts from MIMIC-IV showed superior performance (AUC 0.65–0.76) and identified clinically interpretable predictors.

Key Findings

• TECO achieved AUC 0.89–0.97 for ICU mortality prediction in development COVID-19 cohort and outperformed EDI, RF, and XGBoost.
• External validation in ARDS (n=2,799) and sepsis (n=6,622) cohorts showed higher AUC (0.65–0.76) than RF and XGBoost.
• The model identified clinically interpretable predictors correlated with mortality.

Clinical Implications

If prospectively validated and integrated, TECO could augment early warning systems for ARDS and sepsis, enabling earlier escalation, targeted monitoring, and improved triage. Implementation requires model governance, local recalibration, and bias auditing.

Limitations

• Preprint not yet peer-reviewed; potential overfitting despite external validation
• Generalizability to non-MIMIC settings and prospective performance remain uncertain

Future Directions

Prospective, multi-center impact evaluation; fairness and drift monitoring; integration with clinician-in-the-loop workflows; head-to-head comparison with calibrated risk scores in ARDS.