Explainable machine learning for predicting major adverse limb events following endovascular intervention for peripheral arterial disease

Objective To develop and validate a personalised, explainable machine learning model to predict 30-day major adverse limb events following endovascular intervention for peripheral arterial disease, using a nationally standardised vascular registry dataset, with a focus on identifying modifiable risk factors. Methods A prospectively maintained vascular database aligned to the National Vascular Registry data template, incorporating 1,719 consecutive endovascular interventions for PAD in 1417 patients across five hospitals within two NHS Trusts over five years (2020 to 2024), with all data captured by a dedicated coordinator team independent of the clinical team. The primary outcome was 30-day major adverse limb event (MALE), defined as a composite of acute limb ischaemia, target lesion revascularisation, major lower limb amputation, or limb-related death. Six supervised machine learning algorithms were compared with conservative hyperparameter tuning and bootstrap-derived 95% confidence intervals. Discrimination and calibration were assessed using Area under the receiver operator characteristic curve (AUROC) and Brier score. SHAP-based explainability outputs were structured a priori to identify modifiable risk contributors. Results The primary outcome occurred in 210 of 1,719 procedures (12.2%). On the test set (n=344), Random Forest (AUROC 0.975) and XGBoost (AUROC 0.970) achieved the strongest discrimination. XGBoost was selected as the primary model based on superior calibration (Brier 0.047) and F1 score of 0.753 at a conservative estimate. SHAP analysis identified prior adverse limb events, emergency admission, procedural complexity, and patient age as dominant non-modifiable predictors, while statin prescription and treatment pathway delay metrics emerged as modifiable contributors, with counterfactual modelling predicting a 22.8% reduction in event rate through simultaneous optimisation of all modifiable factors. Conclusion A personalised, explainable machine learning model can accurately predict 30-day MALE following peripheral endovascular intervention using harmonised coordinator-led registry-based data. The pre-specified modifiability framework advances clinical utility beyond risk quantification, identifying actionable peri-procedural targets. External validation and integration into a point-of-care clinical decision support system are underway.