Hamostaseologie 2025; 45(S 01): S38-S39
DOI: 10.1055/s-0044-1801602
Abstracts
Topics
T-06 Diagnostics and laboratory tests

Tackling the implementation hurdle: User-centric validation of a machine-learning decision support tool for the screening of mild bleeding disorders

Authors

  • H Nilius

    1   Inselspital, Bern University Hospital, University of Bern, Department of Clinical Chemistry, Bern, Switzerland
    2   University of Bern, Graduate School for Health Sciences, Bern, Switzerland

  • J Kaufmann

    1   Inselspital, Bern University Hospital, University of Bern, Department of Clinical Chemistry, Bern, Switzerland

  • F Minervini

    3   Cantonal Hospital of Lucerne, Division of Thoracic Surgery, Lucerne, Switzerland

  • M Adler

    4   Inselspital, Bern University Hospital, University of Bern, Department of Hematology and Central Hematology Laboratory, Bern, Switzerland

  • A Greguare-Sander Wieland

    4   Inselspital, Bern University Hospital, University of Bern, Department of Hematology and Central Hematology Laboratory, Bern, Switzerland

  • L Alberio

    5   Lausanne University Hospital CHUV, Division of Hematology and Central Hematology Laboratory, Lausanne, Switzerland

  • D Kröll

    6   Inselspital, Bern University Hospital, University of Bern, Department of Visceral Surgery and Medicine, Bern, Switzerland

  • S Veerakatty

    7   Spital Interlaken, Department of General Surgery, Interlaken, Switzerland

  • B Gerber

    8   Oncology Institute of Southern Switzerland, Clinic of Hematology, Bellinzona, Switzerland

  • G Erdös

    9   Inselspital, Bern University Hospital, University of Bern, Department of Anaesthesiology and Pain Therapy, Bern, Switzerland

  • T Sauter

    10   Inselspital, Bern University Hospital, University of Bern, Department of Emergency Medicine, Bern, Switzerland

  • A Koster

    11   Sana-Herzzentrum Cottbus, Department of Anaesthesiology, Cottbus, Germany

  • J Levy

    12   Duke University School of Medicine, Department of Anesthesiology, Critical Care, and Surgery, Durham, USA

  • M Nagler

    1   Inselspital, Bern University Hospital, University of Bern, Department of Clinical Chemistry, Bern, Switzerland
Weitere Informationen
Auch verfügbar auf
 

Introduction: Due to the existing data infrastructure, the medical laboratory is an ideal candidate to pioneer machine-learning algorithms (MLA) in medicine. In contrast to radiology, however, very few MLAs have been implemented yet. Recently, we developed an easy-to-use MLA for the screening of inherited mild bleeding disorders (MBDs) before surgery, for which there is no sensible diagnostic tool available. This study aimed to overcome the implementation hurdle by performing a user-centric validation of a MLA for the screening for MBD.

Method: Using detailed data from two prospective cohort studies of patients referred with suspected MBD (n=555; n=217), we trained, externally validated, and implemented a diagnostic MLA (https://toradi-hit.dbmr.unibe.ch/mbdcheck/). To assess user-friendliness, we developed a survey platform comprising a demographic questionnaire, four case vignettes, and the system usability scale (SUS), a validated questionnaire for software applications. The survey was sent out to surgeons, anesthesiologists, and hematologists, and various background data were collected.

Results: The final model included the following predictors and performance in the external validation was better than any other diagnostic tool (AUROC: 0.86; 95% CI: 0.81, 0.90): (1) sex, (2) activated partial thromboplastin time, (3) PFA-200 closure time with an epinephrine collagen cartridge, and (4) a simplified bleeding history based on the ISTH bleeding assessment tool (ISTH BAT). Thirty-three surgeons, 29 anesthesiologists, and 24 hematologists participated in the survey; most physicians had 10-14 years of experience in their respective fields (29.0%). The median time needed to fill out the tool was 72 seconds (interquartile range [IQR]: 49.0, 79.5). The median SUS score was 82.5 (IQR: 72.5, 90,>80=Software with excellent usability).

Conclusion: Using the screening of MBD as a case study, we developed and implemented a user-friendly MLA that performed well in external validation. This experience can serve as a starting point for ML applications in a wide range of clinical problems.



Publikationsverlauf

Artikel online veröffentlicht:
13. Februar 2025

© 2025. Thieme. All rights reserved.

Georg Thieme Verlag KG
Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany