SRS105 - Predicting amputation and mortality in patients undergoing first-time lower limb endovascular revascularisation using machine learning models: a racial and gender subgroup analysis

Abstract Introduction Peripheral arterial disease affects >230 million people worldwide, with higher incidence and mortality than that of several cancers. In the UK, approximately 15 000 lower limb endovascular procedures are performed annually, however amputation and mortality rates despite intervention remain high. Discordant RCT results comparing open surgical bypass and endovascular intervention (EI) highlight the potential role of precision medicine. We develop a prognostication machine learning model to predict amputation and mortality at one year in patients undergoing first-time lower limb EI and perform a subgroup analysis to identify patients potentially disadvantaged by these models. Methods Three machine learning models, elastic net (EN), XGBoost, and random forest (RF) were trained to predict the 1-year composite outcome of major amputation or death, using preoperative variables from patients undergoing first-time EI between 2015 and 2021. Models were validated with 5-fold cross-validation, and performance was evaluated using area under the curve (AUC). Variable importance was assessed, and subgroup analyses were conducted. Results Among 803 patients who met inclusion criteria, 192 (24%) experienced the composite outcome. RF demonstrated optimal performance (AUC: 0.74; 95% c.i.: 0.74–0.75). Charlson comorbidity index, Rutherford Score, Haemoglobin and Estimated Glomerular Filtration Rate were most predictive. Subgroup analyses revealed poorer performance in black patients, reflecting key differences in baseline preoperative characteristics. Conclusions This study demonstrates the ability to predict major amputation and mortality at 1-year in patients undergoing EI using routinely recorded preoperative variables. Subgroup analysis highlights the need for equity-focused model development to ensure accurate and inclusive AI models.