AI-Enabled ComPLETE Model for Better Prediction of CLTIRevascularization Outcomes - PROJECT SUMMARY Patients with chronic limb threatening ischemia (CLTI; aka critical limb ischemia, CLI) who undergo revascularization face high rates of readmissions and complications. Prior work has attempted to better understand the drivers of these poor outcomes, with several factors reliably replicated across studies (e.g., various demographics, comorbidities, procedure type, etc.). Despite this, and methodological strengths of these studies such as large sample sizes, a substantial proportion of variance remains unaccounted for, with these predictive models resulting in Area Under the Receiver Operating Characteristic Curves (AUCs) in the .60s. A likely contributor is the data sources used in these studies, which to date have been largely limited to use of state or nationwide registries and/or structured Electronic Health Record (EHR) data. However, one in three post- surgical complications occur after hospital discharge—a time period in which data are largely unaccounted for in the aforementioned sources. Further, physicians who treat CLTI have indicated that there are likely several other factors associated with poor outcomes that are not reliably available in registry or structured EHR data fields, such as caregiver quality and home environment. This project will close these gaps by supplementing EHR and clinical notes data with prospectively collected data from CLTI patients and caregivers after revascularization. Specifically, data sources will include: validated questionnaires from patients, an Artificial Intelligence (AI)-enabled mobile health (mHealth) app using an Ecological Momentary Assessment method to collect daily surveillance data and wound photos from discharge to the first surgical follow-up (14 days post- discharge), and Natural Language Processing (NLP) tools specific to this population to examine clinical notes. These will be merged with data from a large real-world data (RWD) resource, the OneFlorida+ (~19 million patients) Patient-Centered Clinical Research Network (PCORnet) clinical research network. Together, this will yield more comprehensive predictive modeling of readmissions and complications in CLTI revascularization using machine learning (ML) methods, to provide a more complete picture of the factors that contribute to these poor outcomes. Finally, we will begin an implementation mapping approach by working with patients, caregivers, physicians and clinical staff who treat CLTI patients to determine predictors that are most modifiable. Aims include: (1) Leverage a large retrospective cohort to develop NLP tools and determine initial significant predictors of poor outcomes in CLTI revascularization from RWD; (2) Prospectively collect patient-reported and post- discharge data (via an AI-enabled mHealth app) from a cohort of CLTI revascularization patients, and use ML to model these factors with RWD and NLP to predict readmissions and complications; (3) Identify the most feasibly modifiable determinants based on stakeholders’ perspectives to inform future work on intervention development, adaptation, and implementation. This work will inform interventions to improve CLTI outcomes and provide a path to better data modeling and post-discharge support for other clinical conditions.
