Machine learning-based prediction of early neurological deterioration after stroke embolectomy

This study aimed to develop and validate a predictive model integrating multidimensional factors for early neurological deterioration (END) following endovascular thrombectomy (EVT) in patients with acute ischemic stroke (AIS). The goal was to enable individualized END risk prediction, thereby facilitating early clinical decision-making and precision treatment. A total of 370 patients with AIS who underwent EVT at a single institution between January 2023 and January 2024 were retrospectively enrolled. Patients were randomly allocated to a training set ( n = 259) and a validation set ( n = 111) in a 7:3 ratio. In the training set, variables associated with END were initially screened using univariate analysis ( P < 0.05). Subsequently, independent influencing factors were identified using multivariate logistic regression. The Least Absolute Shrinkage and Selection Operator (LASSO) regression was then employed for feature selection. Based on the selected core variables, three machine learning models—Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN)—were constructed. Model performance was assessed using the area under the curve (AUC), calibration curves, and decision curve analysis (DCA). Furthermore, model interpretability was evaluated using SHapley Additive exPlanations (SHAP) values. Independent risk factors for END after EVT were the NIHSS score, preoperative maximum temperature, D-dimer level, NLR, and immediate postoperative systolic blood pressure ( P < 0.05), while ASPECTS served as a protective factor ( P < 0.05). Among the three machine learning models, the RF model demonstrated superior predictive performance, achieving an AUC of 0.887 in the training set and 0.845 in the validation set. Its calibration curve closely approximated the ideal diagonal line, and the Hosmer-Lemeshow test indicated a good fit. Moreover, within the clinically relevant high-risk threshold range of 0–0.6, the RF model provided the greatest net clinical benefit in both cohorts. SHAP analysis identified ASPECTS and NLR as the primary contributors to the model’s predictions, providing insight into the key drivers of the model’s decision-making. The predictive model developed in this study for END after EVT in patients with AIS—particularly the RF model—exhibits satisfactory discrimination, calibration, and clinical utility in internal validation. However, external validation in independent cohorts is needed before clinical implementation.