320 Diabetic foot ulcer (DFU) – Predicting “At-risk” patients using ML analysis

Objectives/Goals: This study aims to evaluate and establish a reliable strategy for identifying at-risk diabetic foot ulcer cases in hospitalized patients utilizing machine learning, enhancing the accuracy of care delivery and facilitating future research on DFU outcomes. “At-risk” is defined as risk for further serious complications, most notably, amputation. Methods/Study Population: Data analyzed for this work started with a previous computable phenotyping study that compared multiple computable phenotype strategies. The dataset created for subsequent ML analysis includes a random sample of 351 hospital admissions during the years 2022–2024. These hospitalizations were manually curated by SME’s, scored and annotated for their inclusion as hospitalized DFU patients, resulting in a “gold standard” dataset for “at-risk” patients. We analyzed six clinical variables hypothesized to be predictive of “at-risk” status (imaging [XRAY, MRI], specialty consultations, and clinical findings [infection, ulcer]). All possible combinations (n=63) were evaluated using precision, recall, F1-score, and accuracy metrics. The dataset contained 42 positive outcomes (12.0% prevalence). Results/Anticipated Results: The findings show, while INFECTION, comprised of two SNOMED codes, is the strongest single predictor, adding ULCER and XRAY provides optimal balance. This combination achieved an F1-score of 72.1% with 93.2% accuracy, 70.5% precision, and 73.8% recall, correctly identifying 31 of 42 positive cases. Adding additional variables beyond this three-feature model yielded diminishing returns with no meaningful improvement in predictive performance. Discussion/Significance of Impact: This research will then be used to proactively identify potential DFU patients for outreach and intervention. A parsimonious three-variable model (XRAY + infection + ulcer) provides optimal predictive accuracy for identifying patients, offering a clinically practical decision support tool with strong discriminative performance.