Machine learning-based prognostic analysis of patients with status epilepticus in the neurological intensive care...

This study aimed to develop and validate machine learning (ML) models for predicting the prognosis of status epilepticus (SE) patients with multisystem complications.

We developed predictive models using six ML algorithms: least absolute shrinkage and selection operator (LASSO) logistic regression, k-nearest neighbors (KNN), support vector machine (SVM), decision tree (DT), random forest (RF), and extreme gradient boosting (XGBoost). We systematically evaluated the prognostic performance of these models against established clinical scores. Specifically, we compared them with the Status Epilepticus Severity Score (STESS), the Encephalitis-Nonconvulsive Status Epilepticus-Diazepam Resistance-Imaging Abnormalities-Tracheal Intubation (ENDIT) score, and the Epidemiology-based Mortality Score in Status Epilepticus (EMSE).

A total of 169 patients with SE were included in this study. In the test dataset, the areas under the curve (AUC) of the models were 0.660 for DT, 0.644 for RF, 0.663 for SVM, 0.689 for KNN, 0.825 for XGBoost, and 0.610 for LASSO logistic regression.The SHAP analysis revealed the top ten predictors contributing to the XGBoost model: hypoalbuminemia, nutritional risk score, age, ventilation duration, NCSE, GCS score, duration of impaired consciousness, creatinine level, APACHE II score, and CCI.

Compared with the other models and scoring systems, XGBoost demonstrated superior predictive performance, suggesting its potential utility for the early identification of high-risk patients and timely clinical intervention. Hypoalbuminemia was identified as the most important prognostic factor, highlighting the critical role of systemic injury in determining adverse outcomes in SE patients treated within the neurocritical care setting.