The atherogenic index of plasma predicts long-term outcomes in patients with severe coronary artery calcification undergoing rotational atherectomy: a machine learning-based cohort study.

backgroundThe atherogenic index of plasma (AIP) is known to be associated with atherosclerotic burden. However, the prognostic value of AIP in patients with severe coronary artery calcification (CAC) undergoing rotational atherectomy (RA) remains unclear. This study aimed to evaluate the relationship between AIP and adverse outcomes in this patient population and to explore relevant risk factors using explainable machine learning methods.

methodsThis study included patients with severe CAC who underwent RA between January 2017 and October 2024, with a median follow-up of 40.55 months. Patients were divided into three groups according to baseline AIP tertiles. The primary endpoints were cardiovascular death or all-cause death; secondary endpoints included non-fatal myocardial infarction, target vessel revascularization, and stroke. Cox regression and restricted cubic splines were used to assess the association between AIP and endpoint events. Kaplan-Meier survival analysis and log-rank tests were employed to compare differences between groups. LASSO regression was used for feature selection, and six machine learning algorithms were applied to construct predictive models for cardiac death. Finally, the SHAP method was used to interpret the model.

resultsIn a cohort of 513 participants (58.28% male), multivariable Cox analysis showed that compared with the lowest AIP tertile, the highest AIP tertile was associated with significantly increased risks of various adverse events: cardiovascular death (HR 2.40, 95% CI 1.40-4.13), all-cause death (HR 1.68, 95% CI 1.13-2.51), non-fatal myocardial infarction (HR 2.27, 95% CI 1.20-4.29), target vessel revascularization (HR 1.74, 95% CI 1.04-2.90), and stroke (HR 1.92, 95% CI 1.00-3.69). Restricted cubic spline analysis indicated a dose-response relationship between AIP and the risk of adverse outcomes. Subgroup analysis suggested that the association between AIP and mortality was stronger in elderly patients, those with cardiac dysfunction, or those with poor glycemic control. Among the six machine learning algorithms, the random forest model demonstrated the best predictive performance for cardiac death (AUC = 0.800). SHAP analysis identified AIP as one of the key features driving the model's predictions. Kaplan-Meier curves revealed that patients in the high-AIP group had worse long-term clinical outcomes.

conclusionAIP was independently associated with adverse outcomes in patients with severe CAC undergoing RA. The integration of this low-cost, readily available biomarker into explainable machine learning frameworks offers a promising avenue for enhancing risk prediction models.