Explainable machine learning for neurocognitive severity clustering and multivariable profiling after acute coronary syndrome

Neurocognitive impairment is a frequent complication of Acute Coronary Syndrome (ACS). Current research often treats cognitive decline as a monolithic outcome, failing to account for the heterogeneous ways in which psychosocial and clinical factors interact to form distinct patient profiles. This study aimed to identify neurocognitive severity clusters in patients following ACS using a data-driven approach and to examine the psychosocial and clinical variables associated with cluster membership. A two-stage machine learning pipeline was implemented in a cohort of ACS patients. In the first stage (unsupervised phase), K-means clustering was used to discover latent neurocognitive clusters. In the second stage (supervised phase), seven machine learning algorithms were evaluated to recover cluster membership from non-cognitive demographic, psychosocial, and clinical variables. Model interpretability was achieved through Partial Dependence Plots (PDP) and probability heat maps. Two distinct clusters were identified: a ' lower-severity cluster' (n = 231) and a 'higher- severity cluster' (n = 100). XGBoost showed highest performance for recovering cluster membership (AUC = .959, Kappa = .713). When designating the severe impairment cluster as the positive clinical class, the model achieved a specificity of 99.1% and a sensitivity of 65.7%. Social support satisfaction was ranked among the most important predictors in six of the seven models. PDP analysis identified a model-derived inflection region between ESSS scores of 35 and 38, while the probability heat map suggested that higher social support was associated with a lower model-estimated probability of membership in the higher-severity cluster, particularly in the presence of elevated depressive symptoms. Explainable machine learning identified neurocognitive severity clusters and highlighted multivariable associations between psychosocial factors and cluster membership following ACS. Social support satisfaction consistently emerged as a highly ranked predictor across multiple algorithms; however, these findings reflect exploratory, model-based associations within a single cohort and require external validation before clinical application.