Integrating frailty and cumulative lipid burden for stroke risk stratification: a machine learning-guided Athero-Frailty Score from the CHARLS cohort.

Abstract

BACKGROUND: Traditional lipid indices, such as the cumulative Atherogenic Index of Plasma (cumAIP), demonstrate inconsistent predictive performance for stroke in older adults, likely due to the modifying effects of complex medication regimens. We hypothesized that incorporating the Frailty Index (FI)—a measure of cumulative physiological deficits—could capture the residual risk missed by lipid markers alone. We developed and evaluated a novel Athero-Frailty Score (AFS) to address these limitations. METHODS: We analyzed 3,690 participants from the China Health and Retirement Longitudinal Study (CHARLS) using a landmark analysis design (baseline: 2015). An exploratory XGBoost model with SHAP analysis was used for variable screening. Based on the orthogonality of lipid and frailty metrics, AFS was constructed as an additive composite score of cumAIP and FI. Associations with incident stroke were assessed using Cox proportional hazards models and restricted cubic splines (RCS). Clinical utility was evaluated via C-statistics, Net Reclassification Improvement (NRI), and Integrated Discrimination Improvement (IDI). RESULTS: SHAP analysis identified FI as having a larger average SHAP contribution than cumAIP and medication indicators. Survival analyses suggested that the association between cumAIP and stroke was attenuated after adjustment for antihypertensive and lipid-lowering therapies (HR 1.20, P = 0.042), and RCS indicated a non-linear pattern with an apparent plateau at higher cumAIP levels. In contrast, AFS was associated with an approximately linear dose–response relationship independent of medication use. In the fully adjusted model, each 1-SD increase in AFS was associated with a higher risk of stroke (HR = 2.75, P < 0.001). Crucially, the addition of AFS yielded significant improvement in reclassification (NRI = 15.4%, P < 0.001), while the improvement in integrated discrimination was modest (IDI = 0.007, P = 0.088). CONCLUSION: The AFS effectively addresses the predictive limitations of traditional metabolic markers in medicated older adults. By integrating metabolic burden with systemic vulnerability, this novel composite score offers a linear and robust predictive approach for stroke, supporting a multidimensional approach to vascular risk stratification in aging populations.