PB1331 - Using natural language processing to diagnose venous thromboembolism: A systematic review

Tuesday, July 12, 2022

6:30 PM – 7:30 PM

Location: Exhibition

Presentation Format: In-Person

Presenting Author(s)

Barbara Lam, MD

Fellow
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States

Co-Author(s)

TC

Thita Chiasakul

Research fellow
Division of Hematology, Department of Medicine, Faculty of Medicine, Chulalongkorn University and King Chulalongkorn Memorial Hospital, Thai Red Cross Society, Bangkok, Thailand
Washington, District of Columbia, United States
DK

Dimitra Karagkouni

Post Doctoral Research Fellow
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
WR

William Robertson

Program Director
National Blood Clot Alliance
Philadelphia, Pennsylvania, United States
Rachel Rosovsky, MD, MPH

MD, MPH. Hematologist
Massachusetts General Hospital and Harvard Medical School
Newton, Massachusetts, United States
LL

Leslie Lake

President
National Blood Clot Alliance
Philadelphia, Pennsylvania, United States
IV

Ioannis Vlachos

Assistant Professor
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
JZ

Jeffrey Zwicker, MD

Chief, Benign Hematology
Beth Israel Deaconess Medical Center
Boston, Massachusetts, United States
RP

Rushad Patell, MD

Assistant Professor of Medicine
Division of Hematology and Hematologic Malignancies, Beth Israel Deaconess Medical Center, Boston, MA
Boston, Massachusetts, United States

Background: Venous thromboembolism (VTE) is the leading cause of preventable in-hospital mortality, and identifying these events is a core metric of hospital quality measures. Understanding the current incidence of VTE can be challenging because organizations often rely on humans to review charts and diagnoses codes. Natural language processing (NLP) offers a promising alternative, allowing for rapid extraction of relevant data from extensive medical records.

Aims: To conduct a systematic review of studies that use NLP to identify VTE events in the electronic health record.

Methods: We conducted a literature search of PubMed, MEDLINE, EMBASE, and Web of Science using Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. Search criteria included original studies with the terms “artificial intelligence” and “venous thromboembolism” published from inception through April 20, 2021. Only studies that employed NLP to assist in the diagnosis or detection of VTE were included. Two reviewers assessed each study for eligibility and adherence to reporting standards using the Transport Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) tool. The pooled sensitivity, specificity, and 95% CI were calculated using a bivariate random effects model.

Results: 556 studies were identified. Sixteen studies with a total of 28 NLP models were included. There was wide variability in adherence to TRIPOD (Figure 1), and some TRIPOD items were deemed not applicable. Sixteen NLP models from 10 studies included sensitivity and specificity as a performance measure. The overall sensitivity of these models was 0.86 (95% CI, 0.73-0.94) and the overall specificity was 0.99 (95% CI, 0.98-0.99) (Figure 2).

Conclusion(s): Sensitivity of NLP to identify VTE events varied greatly between different algorithms, but overall specificity was high. Further studies using standardized performance measures are needed.