Multi-modal prediction of extracorporeal support-a resource intensive therapy, utilizing a large national database.

Summary

Using N3C data (101,400 patients), a hierarchical deep-learning model continuously predicted ECMO utilization up to 96 hours before initiation and outperformed multiple traditional machine-learning baselines at all horizons. Model interpretability highlighted dynamic feature importance across clinical trajectories.

Key Findings

• Developed a hierarchical deep-learning model integrating static and multi-granularity time series features from N3C.
• Included 101,400 patients with 1,298 (1.28%) receiving ECMO support.
• Outperformed Logistic Regression, SVM, Random Forest, and XGBoost in accuracy and precision from 0 to 96 hours before ECMO initiation.
• Interpretability analyses revealed evolving feature contributions across patients’ clinical courses.

Clinical Implications

If prospectively validated, PreEMPT-ECMO could trigger earlier consultation, transfer, or cannulation planning, standardize referral thresholds, and reduce delays for refractory respiratory failure.

Limitations

• Retrospective model development with potential confounding and dataset shift.
• Developed in COVID-era cohorts; generalizability to non-COVID refractory respiratory failure requires prospective validation.

Future Directions

Prospective, multi-center silent deployment with human-in-the-loop evaluation, fairness auditing, and external validation on non-COVID ARDS to assess clinical impact and safety.