VTE Risk Assessment and Management: Machine Learning from Clinical and Genetic Information

The VA Office of Inspector General (OIG) has identified prevention of venous thromboembolism (VTE, i.e., deep vein thrombosis [DVT] and pulmonary embolism [PE]) in hospitalized Veterans as a missed opportunity to improve medical care. VTE affects ~900,000 Americans, with ~100,000 deaths each year. Over half of all VTEs occur after a hospitalization and are preventable, yet only half of hospitalized Veterans receive appropriate VTE prophylaxis. The VA OIG noted that risk stratification, followed by risk-informed prophylaxis of patients at high risk, are the keys to VTE prevention. The ideal solution is a clinical decision support tool that not only assesses a patient's VTE risk; but also evaluates the benefit of VTE prevention vs. the associated risk of bleeding, considers the effectiveness of VTE prophylaxis strategies (none vs. mechanical vs. pharmacologic vs. mechanical + pharmacologic), and accounts for patient preferences regarding prevention of VTE vs. the risk they associate with prophylaxis-associated bleeding. The Centers for Disease Control and Prevention (CDC) has acknowledged the lack of a perfect tool to guide decisions on VTE prophylaxis. There are over 20 risk assessment models (RAMs) for VTE and two for bleeding, all based solely on known clinical risk factors. In a study of >1.2 million Veterans admitted to a VA hospital between 2016 and 2021, we found that these RAMs have poor ability to predict VTE or bleeding, implying that currently used risk factors are insufficient predictors of VTE or bleeding. Genetic predisposition contributes to risk for VTE and for bleeding, and the risk associated with genetic loci may increase with hospitalization. Despite this, RAMs predicting VTE, and bleeding do not include genetic risk factors. They also do not assess separate risk for DVT and PE, even though these conditions have different clinical implications. Moreover, the RAMs are not well validated, are based on limited follow-up, and require labor intensive manual computation. Decisions on prophylaxis resulting from these RAMs are therefore based on the assumed short- term risk for VTE only; not accounting for prophylaxis-associated bleeding risk, effectiveness of prophylaxis, genetic risk factors, or patient preferences. Using the most comprehensive genome wide association study (GWAS) meta-analysis of VTE available, we will generate polygenic risk scores (PRS) for VTE, DVT and PE. We will also conduct a GWAS of bleeding risk in Million Veteran Program (MVP) patients, to generate PRS for bleeding. We will combine genetic data from MVP with clinical data from VINCI (the repository of all data from the VA's EMR) on the ~534,000 MVP participants who were hospitalized, ~15,000 of whom suffered a subsequent VTE and ~30,000 a major bleeding event. We will use Machine Learning (ML) to develop and validate four RAMs predicting 90-day VTE, DVT, PE, and bleeding. The models will include known risks and benefits of VTE prophylaxis strategies and known patient preferences for prevention of VTE vs. risk of bleeding, to develop an automated clinical decision support tool that guides VTE prophylaxis, and which is automatically updated as the patient's condition evolves. The Specific Aims are: Aim 1. Use genetic and clinical data to develop and validate comprehensive RAMs predicting hospitalization related 90-day VTE, DVT alone, and PE with or without DVT. Aim 2. Use genetic and clinical data to develop and validate a comprehensive RAM predicting 90-day major bleeding in hospitalized patients. Aim 3. Develop a decision support tool to guide VTE prophylaxis selection in hospitalized patients that considers 90-day and life-time risk vs. benefit. Our goal is to develop an ML decision support tool that guides VTE prophylaxis in hospitalized patients that is standardized, automated, and updated for daily review. Future work will include integrating the tool into the VA electronic medical record and prospectively validating it in a clinica Project Number: 1I01CX002747-01A1 | Fiscal Year: 2025 | NIH Institute/Center: Veterans Affairs (VA) | Principal Investigator: Brajesh Lal (+1 co-PI) | Institution: BALTIMORE VA MEDICAL CENTER, BALTIMORE, MD | Activity Code: I01 | Study Section: Special Emphasis Panel[ZRD1 CARB-C (01)] View on NIH RePORTER: https://reporter.nih.gov/project-details/11052800