Development and validation of a screening tool for sepsis without laboratory results in the emergency department: a machine learning study.

Summary

Using non-laboratory ED data, the qSepsis machine-learning model (logistic regression) achieved AUROC 0.862 internally and 0.766 in external validation, outperforming SIRS, qSOFA, and MEWS. Precision-recall analysis (AUPRC 0.213) indicates better case finding than comparators but with low PPV that may increase false alarms in practice.

Key Findings

• Logistic regression qSepsis achieved AUROC 0.862 (95% CI 0.855–0.869) internally and 0.766 (0.758–0.774) externally.
• qSepsis outperformed SIRS (AUROC 0.704), qSOFA (0.579), and MEWS (0.600) in external validation.
• AUPRC was 0.213 for qSepsis versus 0.071 (SIRS), 0.096 (qSOFA), and 0.083 (MEWS), but with low PPV indicating potential false alarms.

Clinical Implications

qSepsis can aid rapid triage when labs are delayed or unavailable, especially in prehospital or high-throughput ED settings. It should augment, not replace, clinician judgment, and be paired with confirmatory testing due to low PPV; prospective implementation studies are needed.

Limitations

• Retrospective design with potential labeling bias tied to Sepsis-3 definitions
• Low PPV may increase false alarms; no prospective impact or clinical workflow evaluation

Future Directions

Prospective, multi-site implementation trials to assess clinical impact, alarm burden, and equity; model calibration across diverse EDs; integration with prehospital triage systems.