PB1314 - Risk Assessing Medical Inpatients for Bleeding Events: The Medical Inpatients Thrombosis and Hemostasis (MITH) Study

Tuesday, June 27, 2023

18:30 – 19:30 ET

Room: Exhibition

Presenting Author(s)

Hanny Al-Samkari, MD (he/him/his)

Assistant Professor of Medicine
Harvard Medical School
Boston, Massachusetts, United States

Co-Author(s)

KW

Katherine S. Wilkinson, MS

Statistician
The Larner College of Medicine at the University of Vermont
Burlington, Vermont, United States
NR

Nicholas S. Roetker

Epidemiologist
Hennepin Healthcare Research Institute
Minneapolis, Minnesota, United States
Andrew D. Sparks, MS

Statistician
The Larner College of Medicine at the University of Vermont
Burlington, Vermont, United States
Allen B. Repp, MD, MSc (he/him/his)

Professor
Department of Medicine, The Larner College of Medicine at the University of Vermont
Burlington, Vermont, United States
Ryan Thomas, MD

Hematology Fellow
Hematology/Oncology Division, Department of Medicine, The Larner College of Medicine at the University of Vermont
Essex Junction, Vermont, United States
NS

Nicholas L. Smith

Professor, Epidemiology
Department of Epidemiology, School of Public Health, University of Washington
Seattle, Washington, United States
Chris Holmes, MD PhD

Professor of Medicine
University of Vermont
Burlington, Vermont, United States
RP

Ryan Packer

Biostatistician
University of Vermont
Burlington, Vermont, United States
Mary Cushman, MD, MSc (she/her/hers)

Professor
Department of Medicine, Larner College of Medicine at the University of Vermont, USA
Colchester, Vermont, United States
Neil A. Zakai, MD, MSc (he/him/his)

Professor
Hematology/Oncology Division, Department of Medicine, The Larner College of Medicine at the University of Vermont
Burlington, Vermont, United States

Background: Existing risk assessment models (RAMs) for hospital-acquired bleeding (HA-bleeding) in medical inpatients have limited predictive accuracy and clinical utility.

Aims: To develop RAMs for HA-bleeding overall and by body site.

Methods: Using validated electronic health record (EHR) derived definitions (computable phenotypes), we captured present on admission (POA) and HA-bleeding from all medical admissions between 2010-2019 at the University of Vermont Medical Center. After excluding POA-bleeding, Bayesian logistic regression with a least absolute shrinkage and selection operator (LASSO) prior probability distribution was used to develop a model for HA-bleeding. Variables with a t-statistic ≥1.5 or ≤-1.5 were included in the models. Model performance was assessed using the area under the receiver operating characteristic (AUC) curve and calibration slope.

Results: There were 62,468 medical admissions in the study period; 5,303 admissions were excluded due to POA-bleeding and there were 1992 HA-bleeding events (incidence rate 35 bleeds/1000 admissions). Mean age was 65.4 years (standard deviation, 17.5 years) and 50.9% were male. The model to predict HA-bleeding included 20 risk factors and had modest discrimination (AUC 0.66). Body site-specific bleeding models included different risk factors (Table). Models for certain body sites, e.g. gastrointestinal bleeding, performed better than the model for all bleeding (AUC 0.73 versus 0.66) (Figure, Table).

Conclusion(s): When developing a RAM for HA-bleeding, we determined that each bleeding site had different risk factors. Based on these findings, site-specific bleeding RAMs may be better for use in clinical practice and research than all-site RAMs to improve precision in prediction of bleeding risk.