Predicting Residual Cognitive Deficit Post-Ischemic Stroke: An Imbalance-Aware Machine Learning Pipeline on EHR Data

Residual cognitive deficit (RCD) following ischemic stroke is a critical yet under-investigated aspect of long-term recovery. This exploratory study leverages publicly available, ethnically diverse patient data to assess potential utility and evaluate feasibility of integrating structured clinical, pharmacologic, and laboratory data with machine learning to identify associations with RCD. Two patient cohorts were constructed of ischemic stroke survivors, those with and without documented RCD, identified through standardized SNOMED codes to ensure reproducibility and interoperability. Due to the pronounced class imbalance (59 RCD vs. 566 non-RCD patients), gradient boosting classifiers incorporating Synthetic Minority Over-sampling Technique (SMOTE) and class weighting were employed as part of an imbalance-aware methodological pipeline and to enhance predictive modeling. The gradient boosting model yielded an area under the receiver operating characteristic curve (AUC-ROC) of approximately 0.795, demonstrating promising discriminatory potential. Beyond metrics, this work illustrates the early results and feasibility of combining multi-modal clinical features with reproducible phenotyping standards to advance equity-focused machine learning for healthcare. Given the small number of RCD cases and reliance on synthetic oversampling, the findings are best interpreted as methodological rather than clinical, underscoring the importance of larger externally validated cohorts. Taken together, these early results provide a framework for rare-event prediction in EHR and protected data, highlighting the need for validation in larger, better-balanced cohorts.