Deep Learning Strategies for Predicting Amputation Free Survival in Patients with Peripheral Artery Disease.

Abstract

This study aimed to address the limitations of traditional Cox proportional hazards (CPH) models in predicting amputation free survival in patients with peripheral artery disease (PAD) by benchmarking alternative survival machine learning (ML) models. It evaluated the performance of ML models in capturing non-linear relationships, re-assessing key predictors, and developing a prototype for a patient level risk prediction tool. A retrospective, observational cohort dataset from Waikato Hospital (New Zealand) was analysed, including 2 366 symptomatic patients with PAD who underwent revascularisation between 2010 and 2021. Clinical, biological, procedural information, and outcomes (amputations and deaths) were acquired. The study investigated non-competing risk models (CPH, conditional survival forest, random survival forest, and non-linear CPH [NLCH]) and competing risk models (Fine and Gray subdistribution hazard model and DeepHit model). Models were developed using fivefold cross validation (80/20 training-validation split) stratified by median time to amputation. Performances were evaluated using the concordance index and integrated Brier score. In the fivefold validation sets, the models achieved a mean concordance index of 0.693 (95% confidence interval [CI] 0.674 - 0.711) for the CPH and 0.707 (95% CI 0.664 - 0.750) for conditional survival forest; 0.704 (95% CI 0.667 - 0.741) for random survival forest; 0.704 (95% CI 0.689 - 0.719) for NLCH; 0.619 (95% CI 0.556 - 0.671) for the Fine and Gray model; and 0.667 (95% CI 0.656 - 0.674) for the DeepHit model. Regarding calibration, the NLCH model had a statistically significantly smaller integrated Brier score value of 0.095 (95% CI 0.092 - 0.098). Feature importance analysis identified the main predictors as disease status (chronic limb threatening ischaemia or claudication), diabetes, and medical treatment. A tool incorporating key predictors was developed for patient specific risk stratification. ML models offer advances in predicting amputation free survival in patients with PAD, to further support clinical decision making, although external validation is required before use in clinical practice.