Machine Learning-based prediction of Early Neurological Deterioration after Thrombolysis in Acute Ischemic Stroke

Abstract

ABSTRACTBackgroundEarly neurological deterioration (END) after thrombolysis in acute ischemic stroke (AIS) cannot be ignored. Our aim is to establish an interpretable machine learning (ML) prediction model for clinical practice.MethodsPatients in this study were enrolled from a prospective, multi-center, web-based registry database. Demographic information, treatment information and laboratory tests were collected. END was defined as an increase of ⩾2 points in total National Institutes of Health Stroke Scale (NIHSS) score within 24 hours after thrombolysis. Eight ML models were trained in the training set (70%) and the tuned models were evaluated in the test set (30%) by calculating the area under the curve (AUC), sensitivity, specificity, accuracy, and F1 scores. Calibration curves were plotted and brier scores were calculated. The SHapley Additive exPlanations (SHAP) analysis and web application were developed for interpretation and practice.ResultsA total of 1956 patients were included in the analysis. Of these, 305 patients (15.6%) experienced END. We used logistic regression to identify six important variables: hemoglobin, white blood cell count, the ratio of lymphocytes to monocytes (LMR), thrombin time, onset to treatment time, and prothrombin time. In the test set, the results showed that the Extreme gradient boosting (XGB) model (AUC 0.754, accuracy 0.722, sensitivity 0.723, specificity 0.720, F1 score 0.451) exhibited relatively good performance. Calibration curves showed good agreement between the predicted and true probabilities of the XGB (brier score=0.016) model. We further developed a web application based on it by entering the values of the variables (https://ce-bit123-ml-app1-13tuat.streamlit.app/).ConclusionsThrough the identification of critical features and ML algorithms, we developed a web application to help clinicians identify high-risk of END after thrombolysis in AIS patients more quickly, easily and accurately as well as making timely clinical decisions.