Fast and interpretable mortality risk scores for critical care patients.

Summary

GroupFasterRisk yields sparse, interpretable ICU mortality risk scores that outperform OASIS/SAPS II and match APACHE IV/IVa with a fraction of parameters. In sepsis, AMI, heart failure, and AKI subgroups, it surpassed OASIS and SOFA, and its variable selection improved performance of other ML models.

Key Findings

• GroupFasterRisk outperformed OASIS and SAPS II and matched APACHE IV/IVa using at most one-third of parameters.
• In sepsis/septicemia, AMI, heart failure, and AKI subgroups, GroupFasterRisk beat OASIS and SOFA.
• Variable selection by GroupFasterRisk improved performance of other mortality ML models.
• The algorithm enforces sparsity, group structure, and monotonicity, and yields multiple equally good models.

Clinical Implications

Hospitals can adopt interpretable risk scores with competitive accuracy to guide triage, resource allocation, and sepsis pathways, while enabling clinician oversight and auditing. Variable sets identified by GroupFasterRisk may also refine existing scoring systems.

Limitations

• Retrospective validation; no prospective impact evaluation on patient outcomes.
• Exact sample sizes and external multi-center clinical deployment are not reported.

Future Directions

Prospective multi-center trials to assess clinical impact and fairness, EHR integration, and tailored sepsis-specific score deployment with clinician-in-the-loop refinement.