Interpretable machine learning prediction model for major adverse cardiovascular events in patients with peripheral artery disease.

BACKGROUND

Major adverse cardiovascular events (MACE) are severe complications of peripheral arterial disease (PAD), associated with a poor prognosis and disease burden. Therefore, the early identification of high-risk individuals is of paramount importance. This study aimed to develop and validate an interpretable machine learning (ML)-based prediction model for MACE risk in patients with PAD.

METHODS

This retrospective study included patients with PAD enrolled between January 2022 and December 2023, with follow-up had been completed by December 2024. The primary outcome was MACE, defined as a composite of myocardial infarction, stroke, and cardiovascular mortality, and patients were followed up for 12-24 months using nonoverlapping datasets from four centers: three for model training and internal validation and one for external validation. Feature selection was performed using univariate analysis, least absolute shrinkage and selection operator logistic regression, and a random Forest algorithm. Ten different ML algorithms were used to construct the risk prediction model. Model performance was evaluated based on discrimination and calibration. The SHapley Additive exPlanations method was used to visualize model features and individual case predictions. The final risk prediction model was presented as a web-based calculator.

RESULTS

This multicenter study involved both model development dataset (n = 1110) and external validation dataset (n = 448). Among the 1558 enrolled patients with PAD, 469 of 1558 patients (30.1%) experienced MACE. The incidence of MACE was higher in the training cohort (249/777 [32.0%]) compared with the internal validation cohort (102/333 [30.6%]) and external validation cohort (118/448 [26.3%]). The mean follow-up duration was 19.0 ± 11.3 months. Participants' mean age was 73.1 ± 10.8 years, with males comprising 70.0% of the patients (1091/1558). We developed ML models incorporating eight clinically significant variables, with Gradient Boosting demonstrating comparatively better performance by achieving area under the receiver operating characteristic curve values of 0.864 (95% confidence interval, 0.822-0.905) in internal validation cohort and 0.777 (95% confidence interval, 0.720-0.833) in external validation cohort. The key predictors included polyvascular disease, cerebrovascular disease, hemoglobin A1c, C-reactive protein, albumin, peripheral arterial surgery, coronary heart disease, and neutrophils.

CONCLUSIONS

The Gradient Boosting algorithm outperformed other models in predicting MACE risk in patients with PAD, with external validation confirming its clinical applicability. The SHapley Additive exPlanations framework and web-based calculator enhanced the model's interpretability, enabling clinicians to better understand the factors contributing to MACE. This tool potentially helps clinicians to identify MACE risk in patients with PAD and implement preventive measures more effectively.