PB0490 - Does a machine learning based algorithm improve dosing of Vitamin K antagonists compared to current decision support programs?

Sunday, June 25, 2023

18:30 – 19:30 ET

Room: Exhibition

Presenting Author(s)

Chantal Visser, MD

PhD candidate, MD
Department of Hematology, Erasmus MC, Erasmus University Medical Center Rotterdam, Rotterdam, The Netherlands
Rotterdam, Zuid-Holland, Netherlands

Co-Author(s)

MH

Mies Huizenga-Dronkert

Dosing physician, MD, PhD
Anticoagulation clinic Star-shl, Rotterdam, The Netherlands
Rotterdam, Zuid-Holland, Netherlands
JZ

Joost J. Zeeuw

Product Manager, PhD
Pacmed, Amsterdam, the Netherlands
Amsterdam, Noord-Holland, Netherlands
AD

Ann Dierkx

CEO
Asolutions, Capelle aan de IJssel, the Netherlands
Capelle a/d IJssel, Zuid-Holland, Netherlands
MB

Maarten Beinema

Dosing Physician, MD, PhD
Anticoagualation Center, Deventer Hospital, Deventer, The Netherlands
Deventer, Overijssel, Netherlands
JB

Jan Paul Barends

Scientific Programmer
Anticoagulation clinic Star-shl, Rotterdam, The Netherlands
Rotterdam, Zuid-Holland, Netherlands
MS

Marieke Schaijk, van

Innoviation Specialist
Anticoagulation clinic Star-shl, Rotterdam, The Netherlands
Rotterdam, Zuid-Holland, Netherlands
MK

Marieke J.H.A. Kruip

Associate professor, hematologist, MD, PhD
Department of Hematology, Erasmus MC, Erasmus University Medical Center Rotterdam, Rotterdam, The Netherlands
Rotterdam, Zuid-Holland, Netherlands

Background: Vitamin K antagonists (VKA) dosing can potentially be improved by computerized decision support programs (CDS), but their impact in out-patient settings has been inconsistent. A machine learning based CDS might yield better results.

Aims: To compare the accuracy of a newly-developed machine learning (ML) algorithm to current CDS for acenocoumarol dosing.

Methods: A ML-based CDS was created to predict the VKA dosage given by the 10% best-performing dosing consultants. We randomly selected 280 INR measurements from 2018, when the Trodis algorithm was used as CDS, and used the Trodis, Beinema II and ML algorithms to generate VKA dosage proposals. The given VKA dosage in 2018 and the algorithms’ proposals were randomly given to six blinded consultants to use as decision support to prescribe the next dosage. We used the 2018 VKA dosage as reference to compare the algorithms’ proposals and the resulting prescription as reference to compare the resulting prescriptions from the algorithms.

Results: The ML proposals consisted of significantly less tablets and the Trodis proposals of more tablets than the reference (Figure 1). Prescriptions based on the proposals of the algorithms were similar to the prescriptions based on the reference (Figure 1). However, the ML and Trodis proposals were significantly altered by the consultants (Figure 2). In unstable patients, in which Trodis does not generate a proposal, the ML’s proposals and prescriptions were significantly lower, while the Beinema II’s proposals and prescriptions were comparable to the reference and its resulting prescriptions (Figure 1).

Conclusion(s): The Beinema II algorithm’s proposals and resulting prescriptions were most similar to the reference and its resulting prescription, while the ML algorithm’s proposals were significantly different from the reference and altered by the consultants. These results suggest that the development of a CDS that utilizes machine learning for the dosing of VKAs is challenging.