The social support "Staircase": Decoding neurocognitive phenotypes in acute coronary syndrome via explainable machine learning

Abstract

Background: Despite the high prevalence of neurocognitive impairment following Acute Coronary Syndrome (ACS) [1], clinical management often overlooks the heterogeneous nature of these deficits [2]. Identifying distinct neurocognitive phenotypes is essential for personalized rehabilitation [3]. Objective: To organize post-ACS neurocognitive profiles using a data-driven pipeline and determine the non-linear predictors of severe impairment. Methods: We applied a two-stage machine learning framework to an ACS cohort. First, an unsupervised phase (K-means clustering) was used to discover latent phenotypes based on cognitive performance. Second, a supervised phase compared seven machine learning algorithms to predict phenotype membership. Explainable AI (XAI) tools, including Partial Dependence Plots (PDPs) and probability heat maps, were used to visualize variable interactions. Results: Two phenotypes emerged: "Mild/Moderate" (n=231) and "Severe Impairment" (n = 100). XGBoost outperformed all other models (AUC = 0.959; Sensitivity = 99.1%). A robust algorithmic consensus was achieved, with six of the seven models identifying Social Support (ESSS) as the primary predictor. XAI analysis revealed a critical "staircase effect": neurocognitive risk remains high and stagnant until a social support threshold of 35–38 points is reached. Furthermore, high social support was found to exert a "buffering effect", significantly neutralizing the cognitive impact of high depressive symptoms. Conclusions: Neurocognitive health post-ACS is not a linear function of clinical severity but a complex interplay of psychosocial resources. The identification of a specific social support threshold (ESSS < 35) provides a concrete clinical marker for identifying patients at risk of severe decline, necessitating a shift toward socially integrated cardiac rehabilitation.