Comprehensive Summary
This research investigates the possibility for a machine-learning model to predict post-stroke cognitive impairment (PSCI) in patients 3–6 months after acute ischemic stroke. The authors recruited 494 patients with stroke, derived 49 admission clinical, imaging, laboratory, and scale variables, and split the cohort into training (70 %) and validation sets (30 %). They applied LASSO and logistic regression for feature selection, and then built and compared seven machine learning models (e.g., XGBoost, logistic regression, SVM, etc.), cross-validation and evaluation metrics (AUROC, sensitivity, specificity, etc.). Their best model, XGBoost, produced an AUROC of 0.980 on training and 0.887 on validation. The most significant predictors were depression score (HAMD-24), sleep quality (PSQI), albumin level, fasting blood glucose, age, NIHSS, hypertension, number of lesions, and presence of paraventricular lesion. They write in the article that the tool may help clinicians to early stratify risk of PSCI and that interpretability via SHAP allows understanding of how certain features affect prediction, but indicate caveats as single-center data and need for external validation.
Outcomes and Implications
This research is significant since PSCI is disabling and widespread, but early identification of high-risk patients is challenging. If, based on data available at admission, clinicians could have forecast post-stroke cognitive decline risk, they would have been able to target interventions more effectively. In practice, the model (especially the XGBoost variant augmented with a nomogram) could guide personalized follow-up, rehabilitation planning, or prevention (e.g. depression therapy, sleep improvement, glucose or nutrition management). But clinical use isn't instantaneous because there would need to be multicenter external validation and prospective trials to show that responding to predicted risk makes a difference in outcome.