Development, validation, and user-centric evaluation of an interpretable machine learning decision support tool for the preoperative prediction of mild bleeding disorders (MBD-Check): a prospective diagnostic prediction study.

Summary

An interpretable ML tool (MBD-Check) using activated partial thromboplastin time, epinephrine-collagen PFA, sex, and a streamlined bleeding history achieved AUROC 0.85 in external validation, with 90.2% sensitivity and 54.3% specificity. Median completion time was 72 seconds and usability was excellent (median SUS 82.5), supporting real-world deployment to streamline preoperative referrals.

Key Findings

• Selected predictors were activated partial thromboplastin time, PFA (epinephrine-collagen), sex, and a streamlined bleeding history.
• External validation showed AUROC 0.85 with 90.2% sensitivity and 54.3% specificity.
• Usability testing across surgeons, anesthesiologists, and hematologists showed a median SUS score of 82.5 and a 72-second median completion time.

Clinical Implications

Integrate MBD-Check into preoperative workflows to identify high-sensitivity candidates for hematology referral while minimizing over-referral. It uses readily available tests, facilitating implementation without new infrastructure.

Limitations

• Developed and validated in two Swiss centers; generalizability to broader, multi-national settings needs confirmation
• Moderate specificity may still lead to some over-referral; impact on outcomes and costs requires prospective implementation studies

Future Directions

Multicenter implementation trials to assess impact on diagnostic yield, perioperative bleeding outcomes, and cost-effectiveness; calibration in diverse populations; integration with EHR and reflex testing.