Artificial Intelligence-Based ABI Dynamic Fluctuation Patterns Predict Adverse Vascular Events in PAD: A Multicenter Prospective Study.

Abstract

Objective: To develop an artificial intelligence-based predictive model utilizing ankle-brachial index (ABI) dynamic fluctuation patterns and evaluate its predictive value for major adverse limb events (MALE) in patients with peripheral arterial disease (PAD), thereby providing a novel risk stratification tool for precision medicine.

Methods: This multicenter prospective cohort study enrolled 412 consecutive PAD patients from six tertiary hospitals between January 2020 and December 2022. The cohort included 289 males (70.1%) with a mean age of 67.8±12.1 years. Using a standardized ABI measurement protocol, values were obtained at baseline, 1, 3, 6, 12, 18, and 24 months. The ABI dynamic fluctuation index (ABI-DFI) was defined as a composite metric incorporating the standardized ABI coefficient of variation with trend analysis. Machine learning algorithms (random forest, support vector machine, neural network) were employed to construct MALE prediction models. The primary outcome was MALE within 24 months, including major amputation and failed revascularization. Major adverse cardiovascular events (MACE), including cardiovascular death, were analyzed separately. Time-dependent ROC curves, competing risk models, and nomogram approaches were utilized to assess predictive performance.

Results: During the 24-month follow-up, 73 patients (17.7%) experienced MALE, including major amputation in 31 cases (7.5%) and failed revascularization in 42 cases (10.2%). Additionally, 16 patients (3.9%) experienced cardiovascular death, which was analyzed as part of MACE. The ABI-DFI was significantly higher in the MALE group compared to the non-MALE group (0.34±0.12 vs. 0.18±0.08, P<0.001). The random forest algorithm-based prediction model demonstrated superior performance with a time-dependent area under the ROC curve (td-AUC) of 0.847 (95%CI: 0.801-0.893) in the validation dataset and 0.831 (95%CI: 0.789-0.873) in the testing dataset, significantly outperforming traditional single-point ABI values (td-AUC=0.692, P<0.001). The optimal ABI-DFI cut-off value was 0.26, with sensitivity of 81.5% and specificity of 78.2% in the validation dataset. When applied to the independent testing dataset, this cut-off demonstrated a sensitivity of 79.8% and specificity of 76.4%. After adjusting for traditional risk factors, the competing risk model identified ABI-DFI as an independent predictor of MALE (HR=3.42, 95%CI: 2.18-5.37, P<0.001). The nomogram prediction model exhibited a C-index of 0.834, with bootstrap validation demonstrating good calibration and discriminative capability.

Conclusion: The artificial intelligence-based ABI dynamic fluctuation prediction model accurately predicts the risk of adverse vascular events in PAD patients, offering significant advantages over traditional assessment methods and providing new scientific evidence for clinical precision medicine and individualized treatment decisions.