A predictive model for early neurological deterioration in medullary infarction based on explainable machine learning.

Abstract

Background: Medullary infarction is a severe subtype of ischemic stroke. Early neurological deterioration (END) is a common adverse event that can significantly affect clinical outcomes.

Objective: This study aimed to develop a machine learning model to effectively predict the risk of END in patients with medullary infarction, thereby providing a reference for clinical intervention.

Methods: In this multicenter retrospective study, we enrolled patients with acute medullary infarction who received treatment at three hospitals located in different regions between 2019 and October 2024. We collected comprehensive, multi-dimensional clinical data and employed five distinct algorithms to construct predictive models. The performance of these models was thoroughly evaluated using accuracy, area under the receiver operating characteristic curve (AUC), precision-recall curves, and calibration curves.

Results: A total of 352 patients were enrolled in this study. The incidence of END was 10% (35), while the incidence of non-END was 90% (317). Seven features were used to build prediction models with the five algorithms. The single-layer neural network demonstrated superior performance in predicting END, achieving an AUC of the receiver operating characteristic (ROC) curve of 0.873 on the validation set, while also exhibiting favorable calibration and a balanced overall performance profile. In comparison, the Naïve Bayes model showed excellent classification consistency and may serve as a complementary tool in clinical practice.

Conclusion: This study developed a highly accurate single-layer neural network and a robust Naïve Bayes classifier, which together provide an interpretable tool for assessing the risk of END in stroke patients across various clinical settings.