AI-driven Clinical Decision Support to Reduce Hospital-Acquired Venous Thromboembolism: Study Protocol for the VTE-AI Randomized Trial.

Hospital-acquired blood clots (HA-VTE) are the leading cause of death in hospitalized patients in the US. Each year, about 900,000 people get blood clots, costing between $7 and $10 billion in medical expenses. HA-VTE is the second leading cause of long-term disability and causes significant health issues and deaths in both adults and children. About 1 in 3 people who get blood clots experience long-term complications. Reducing HA-VTE is a major challenge. This study will test a new AI method to predict and prevent HA-VTE. The goal is to see if this AI tool can reduce the number of HA-VTE cases in the Vanderbilt Health System, which includes both urban and rural hospitals. The AI tool, called VTE-AI, calculates a risk score without needing input from doctors. It will suggest reconsidering blood clot prevention measures for patients who don't have them ordered and have no reasons to avoid them. This suggestion will be made after admission and daily during the hospital stay. Currently, doctors manually calculate a risk score and choose a prevention option. This study will compare the effectiveness of the AI tool against the current manual method in reducing HA-VTE cases. The study will randomly assign half of the patients to use the AI tool and the other half to the standard manual method.

Background Hospital Acquired Venous Thromboembolism (HA-VTE) remains the leading cause of death in hospitalized patients in the US. Approximately 900,000 people experience VTE each year, with incidence-based medical costs estimated between $7 and $10 billion per year. The second leading cause of disability-adjusted life-years, HA-VTE causes significant morbidity and mortality in adult and pediatric patients. Roughly 1 in 3 people experience long-term complications (i.e., post-thrombotic syndrome) following VTE. Reducing HA-VTE presents a major diagnostic challenge. Despite numerous published prognostic models of HA-VTE, no single model outperforms the rest. And HA-VTE affects groups inequitably, which means models might reflect or worsen healthcare disparities if they are not deployed in the context of responsible, algorithmovigilant systems. Integrating scalable AI for HA-VTE prevention into effective clinical decision support (CDS) might effectively reduce HA-VTE incidence while aiding the realization of the potential for AI in high-value clinical practice. Recently, a Vanderbilt team of clinicians and biostatisticians validated a regression risk score called "VTE-AI" to prognosticate risk of HA-VTE on admission. The urban-rural divide has long caused healthcare disparities in morbidity and mortality. These differences might not result from rurality itself, but from "the effects of socio-economic disadvantage, ethnicity, poorer service availability, higher levels of personal risk and more hazardous environmental, occupational and transportation conditions." AI implementation will be no different without close attention to differences in both deployment settings. Studying multiple simultaneous implementations of AI in both urban and rural setting with adult and pediatric patients will yield unprecedented insights for AI-driven CDS. Rationale and Specific Aims This study will rigorously study a novel AI approach to predicting risk of HA-VTE and guiding prevention. This trial will produce strong evidence for the pragmatic use of novel AI-CDS to prevent HA-VTE across diverse sites and populations. The goal of this study is to evaluate the effectiveness of AI-driven CDS to reduce the incidence of HA-VTE across the Vanderbilt Health System including urban and rural sites: Vanderbilt Adult Hospital (VUH) and the Vanderbilt Regional Health System (VRHS) including Vanderbilt Tullahoma Harton Hospital (VTHH), Vanderbilt Bedford County Hospital (VBCH), and Vanderbilt Wilson County Hospital (VWCH). The investigators will implement a validated risk score, VTE-AI,5 which does not require clinician input to calculate, to prompt CDS suggesting reconsideration of DVT prophylaxis in those who 1) do not have active prophylaxis ordered and 2) have no contraindication to pharmacologic prophylaxis. This CDS "nudge" will occur after admission orders have been submitted for hospital admissions and on each subsequent day of an inpatient encounter. The current standard of care is an order set requiring manual calculation of the Padua risk tool and selection of a prophylaxis option or documentation of a temporary or permanent exception. The investigators will evaluate the effectiveness of the VTE-AI-driven clinical decision support (CDS) against the standard DVT/VTE prophylaxis order set to reduce HA-VTE incidence across urban/rural dimensions. The investigators will conduct a pragmatic RCT of VTE-AI-driven CDS randomizing half the eligible encounters to CDS and half to standard of care. Objectives Primary Objective The investigators hypothesize CDS will reduce incidence of HA-VTE in those i) predicted at high risk by VTE-AI and ii) without evidence of pharmacologic prophylaxis in half, from baseline 4.3% incidence (562/12,946 events) to 2.2% incidence, which will require 2,236 encounters. Sample size calculation indicates at least 1,118 patient encounters are needed in each arm to achieve 80% power with 5% probability of type I error. Using historical data from 2023-2024, there were 150-230 encounters per month meeting these criteria. The investigators will therefore conduct the RCT for one (1) year to meet required sample size. Secondary Objective The investigators hypothesize the VTE-AI CDS will not increase bleeding risk, readmission rates, or lengths of stay (LOS) between the intervention and non-intervention arms. Trial design Two-arm RCT with randomization at encounter-level.