Machine Learning-Based Model for Prediction of Post-Stroke Cognitive Impairment in Acute Ischemic Stroke: A Cross-Sectional Study

Early identification of post-stroke cognitive impairment (PSCI) is an important challenge for clinicians. In this study, we aimed to build a machine learning-based prediction model for PSCI and uncover potential risk factors to support clinical decision-making.

We collected features of 96 patients with acute ischemic stroke and measured cognitive impairment using the Mini-Mental State Examination. Three common machine learning algorithms, including support vector machine, Gaussian naive Bayes, and logistic regression, were used to build clinical prediction models for PSCI. The area under the receiver operating characteristic curve (AUROC), specificity, sensitivity, negative prediction value, positive prediction value, accuracy, and model fitting effect were used to evaluate the predictive performance of the models and further determine the clinical prediction rules.

In this study, the logistic regression model showed the best performance with an AUROC of 0.86, which was higher than the values of the other two models. Moreover, the logistic regression model showed high sensitivity (0.82), specificity (0.83), negative prediction value (0.88), positive prediction value (0.75), and accuracy (0.83). This work identified the top nine factors in importance ranking as predictors of PSCI. Among them, age and urine glucose were significantly associated with PSCI (P < 0.05).

Machine learning algorithms may be useful in the prediction of PSCI, especially logistic regression algorithms. In the present study, aging and hyperglycemia were independent risk factors for PSCI, and the cognition of such patients should be carefully addressed in clinical practice screening work.