P-008 Laboratory medicine Oral communication

When the LIS runs the algorithm: an implementation study of a machine-learning model for heparin-induced thrombocytopenia (RADI-HIT)

Audience Choice Award

To rate this poster, scan the QR code on the poster itself — voting only works on site. Every attendee has 3 stars and can give at most one per poster. A star is final and cannot be taken back.

Authors

H. Nilius presenting

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

#0
M. Stickel

Medical Directorate, Inselspital, Bern University Hospital, University of Bern, Bern

#1
G. Speierer

Medical Directorate, Inselspital, Bern University Hospital, University of Bern, Bern

#2
J. Miazza

Department of Cardiac Surgery, Inselspital, Bern University Hospital, University of Bern, Bern

#3
J.A. Kremer Hovinga

Department of Hematology and Central Hematological Laboratory, Inselspital, Bern University Hospital, University of Bern, Bern

#4
M. Reusser

Directorat of Technology and Innovation, Inselspital, Bern University Hospital, Bern

#5
J. Waskowski

Department of Intensive Care Medicine, Inselspital, Bern University Hospital, University of Bern, Bern

#6
T. Hochgruber

Department of Intensive Care Medicine, Inselspital, Bern University Hospital, University of Bern, Bern

#7
M. Siepe

Department of Cardiac Surgery, Inselspital, Bern University Hospital, University of Bern, Bern

#8
J.C. Schefold

Department of Intensive Care Medicine, Inselspital, Bern University Hospital, University of Bern, Bern

#9
M. Nagler

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

#10

Abstract

Introduction: Confirmatory testing for heparin-induced thrombocytopenia (HIT) takes several days, which can delay treatment decisions. To address this, we previously developed TORADI-HIT, a machine-learning decision support tool combining an anti-PF4/heparin immunoassay, routine lab parameters, and clinical variables. In external validation, it reached high accuracy and is available as a web application (https://toradi-hit.org). In the follow-up study (RADI-HIT), we aimed to embed the algorithm into the laboratory information system (LIS, Epic Beaker) at Inselspital, Bern University Hospital, so it can be applied at the bedside.

Methods: When HIT is suspected, physicians order a defined HIT profile and three clinical variables are entered (two 4Ts items and unfractionated heparin exposure). The LIS checks whether platelet count and CRP from the last 24 h are available; if not, fresh samples are ordered automatically. Once results are validated, TORADI-HIT runs and returns a classification probability with interpretation to the lab result interface. A positive result triggers a pop-up recommending hematology consultation. Standard procedures, including interpretation by laboratory hematologists, run in parallel during the validation phase. Since Epic’s ML module (Nebula) cannot pass results into the LIS for technical reasons, we replaced the original support-vector machine (SVM) with a logistic regression model of comparable performance computed directly in the LIS. The model runs on hospital infrastructure.

Results: Fifty-six suspected HIT cases were included. Median turnaround time was 1.47 h (IQR 1.18–1.9) for TORADI-HIT versus 25.1 h (IQR 11.7–62.3) for hematology interpretation. Five false positives occurred, mostly in patients with high 4Ts items and weakly positive anti-PF4/heparin antibodies. Generally, high acceptance of the integrated workflow was observed.

Conclusion: LIS-embedded machine learning is an underexplored but promising route for clinical AI implementation. It lives inside a workflow physicians already use daily and might remove the friction of separate applications. In RADI-HIT, this approach cut HIT turnaround time by about 17-fold without missing cases. The main constraint was technical: current Epic functionality does not support the return of ML results into the LIS, which is why a switch from SVM to logistic regression was necessary.