IFT04 - Deep Learning Strategies for Predicting Amputation Risk in Patients with Peripheral Artery Disease

Objectives: Peripheral artery disease (PAD) is associated with high morbidity rates and risks of amputation. Although some risk factors of amputation have been identified, the outcomes of patients after revascularization remain difficult to predict. The aim of this study was to apply different ML models to predict amputation in patients who underwent revascularization for PAD and to propose an end-to-end tool for patient-level risk prediction.

Methods: A dataset from a retrospective observational cohort study was used (Waikato Hospital, New Zealand) and included PAD patients who underwent revascularization (from January 1, 2010, to December 31, 2021). Data included clinical, biological, procedural characteristics and outcomes of patients including amputations and deaths. The models tested included the Cox Proportional Hazards model (CPH), Random Survival Forest (RSF), Conditional Survival Forest (CSF) and Non-Linear Proportional Hazard model (NLCH). The benchmark was conducted using 5-fold cross-validation with an 80/20 training-to-testing split stratified by the median time before amputation. The performances of the models were assessed using C-index and Brier score. Predicted survival curves by CPH, RSF, NLCH, and CSF on the first fold test set were compared with the survival curve by Kaplan–Meier (KM). The differences between KM and predicted survival curves were quantified with the root mean square error (RMSE).

Results: In total, 2 366 patients were included and the median follow-up time was 767 ± 462 days. Similar performance values in both C-index and Brier score were observed for the 4 ML models, reaching a C-index of 0.71±0.02 and a Brier score of 0.11±0.021 on the 5-fold test sets to predict amputation-free survival (Table 1). Analysis of feature importance highlighted age, clinical stage, antithrombotic drug, type of surgery, creatinine level, Chronic obstructive pulmonary disease (COPD), and ischemic heart disease as main predictors in the model. The 4 models demonstrated good performances as compared to KM, although the accuracy of the models tended to decrease over time (Figure 1). Overall, the NLCH model reached the lower RMSE (=0.0049) suggesting that the model is closer to the observed risk probability across time. An online tool integrating the main predictor specifications was then built to predict a risk per patient and classify patients into low, medium, and high risk (Figure 2).

Conclusions: This study highlights the interest of ML models to develop patients’ tailored risk prediction of amputation. Such application could help to better evaluate the prognosis of patients and guide therapeutic decision making.