Complete Blood Count and Monocyte Distribution Width-Based Machine Learning Algorithms for Sepsis Detection: Multicentric Development and External Validation Study.

Summary

Using six cohorts (n=5344) from five hospitals, the authors developed ML models combining CBC parameters and MDW for early sepsis detection. Models achieved AUCs of 0.91–0.98 internally and 0.75–0.95 on five external cohorts, outperforming standalone biomarkers and prior ML baselines. Controllable AI features (cautious classification/abstention) and explainable AI improved robustness under label, covariate, and missing-data shifts and yielded interpretable diagnostic rules.

Key Findings

• Developed ML models using CBC and MDW across six cohorts (n=5344) with internal AUC 0.91–0.98 and external AUC 0.75–0.95.
• Outperformed baseline biomarkers and state-of-the-art ML models for sepsis detection.
• Controllable AI (cautious classification/abstention) and explainable AI improved performance and yielded interpretable diagnostic rules under distribution shifts.

Clinical Implications

Hospitals could integrate abstaining, explainable AI models into lab information systems to flag high-risk patients earlier than clinical recognition, while minimizing false positives by allowing abstention under uncertainty.

Limitations

• Retrospective observational datasets; no prospective impact evaluation or randomized deployment.
• Generalizability beyond Italian hospitals remains to be demonstrated; potential spectrum and verification biases.

Future Directions

Prospective, multi-country impact evaluations (e.g., stepped-wedge trials) to assess mortality, time-to-antibiotics, and alarm burden; calibration across lab platforms; integration with EHR triage.