[Establishment and evaluation of a predictive model for early neurological deterioration after intravenous...

To establish a machine learning model to predict the risk of early neurological deterioration (END) based on the clinical and laboratory data of patients with acute ischemic stroke (AIS) before intravenous thrombolysis.

The clinical data of AIS patients who received intravenous thrombolytic with recombinant tissue plasminogen activator (rt-PA) at the Stroke Center of the First Hospital of Qinhuangdao City from January 2019 to July 2022 were retrospectively analyzed. Patients were divided into END group and non-END group according to whether END appeared after intravenous thrombolytic. Clinical data of patients at admission were collected, including demographic characteristics, clinical evaluation, comorbidification, drug use history, laboratory tests, etc. Univariate and multivariate Logistic regression analysis were performed to screen out the independent predictors of the END of AIS patients after intravenous thrombolytic. The study subjects were randomly divided into a training set and a test set in a 7 : 3 ratio. Four machine learning prediction models, including Logistic regression (LR), K-nearest neighbor (KNN), support vector machine (SVM) and random forest (RF), were established based on independent predictors. The receiver operator characteristic curve (ROC curve) was used to evaluate the predictive ability of each model in END.

A total of 704 patients were enrolled, of whom 99 were identified as END and 605 as non-END. Univariate and multivariate Logistic regression analysis was used to screen out the National Institutes of Health stroke scale [NIHSS, odds ratio (OR) = 1.049, 95% confidence interval (95%CI) was 1.015-1.082, P = 0.004], systolic blood pressure (OR = 1.013, 95%CI was 1.004-1.022, P = 0.004), lymphocyte percentage (LYM%, OR = 0.903, 95%CI was 0.853-0.953, P < 0.001), platelet to lymphocyte ratio (PLR, OR = 1.007, 95%CI was 1.002-1.014, P = 0.013) were the independent predictors of END in AIS patients after intravenous thrombolysis. The area under the curve (AUC) of LR, KNN, SVM, and RF machine learning models in the test dataset were 0.789 (95%CI was 0.675-0.902), 0.797 (95%CI was 0.685-0.910), 0.851 (95%CI was 0.751-0.952) and 0.809 (95%CI was 0.699-0.919), respectively. The RF model had the highest sensitivity (95.7%). The accuracy (0.736), specificity (72.0%) and AUC of SVM model were the highest, and its overall prediction ability was better than the other three models.

Machine learning models have a potential role in early predicting the risk of END after intravenous thrombolysis in AIS patients, and can provide help in clinical decision-making for intravenous thrombolysis.