PCRF Abstracts - Details View

 

Download Abstract Poster (PDF)

 

Developing a Machine Learning–Based Prediction Tool for Stroke at the Prehospital Level

Author: Minwoo Kim, Sang Do Shin, Young Sun Ro, and Jeong Ho Park | |

Associate Authors:

Introduction:

Stroke is a significant health problem and one of the leading causes of death and disability. It is critical to transport the patient with suspicion of stroke to the appropriate hospital so that can receive the necessary treatment in a timely manner.

Objective:

To develop a machine learning–based decision-making model to facilitate EMS decision-making by predicting the type of stroke and assessing the necessity of intra-arterial thrombolysis based on prehospital information.

Methods:

A retrospective study was conducted in eight emergency departments (EDs) in South Korea from July 1, 2020, to December 31, 2022. Adult patients who exhibited cerebrovascular disease symptoms in the absence of trauma or cardiac arrest who were transported to an ED by emergency medical services (EMS) were enrolled. The primary outcome of the study was diagnosis of isolated ischemic stroke, subarachnoid hemorrhage, or the other intracranial hemorrhage at the time of ED discharge or hospital admission via the ED. The secondary outcome of the study was whether intra-arterial thrombolysis for ischemic stroke was performed. To predict the outcomes, we developed models using machine learning algorithms: logistic regression analyses (LR) and extreme gradient boost (XGB). We randomly selected 80% of the patient data to develop the model; the other 20% was used as test data.

Results:

A total of 26,824 patients suspected of having cerebrovascular disease in the prehospital environment were enrolled. The area under the receiver operating characteristic curve (AUROC) of the outcomes was 0.8875–0.8994 for isolated ischemic stroke, 0.7920–0.8546 for subarachnoid hemorrhage, and 0.8455–0.8659 for other intracranial hemorrhage. The AUROC of the prediction model for intra-arterial thrombolysis in ischemic stroke was 0.8909–0.9111. All models showed excellent performance in all outcomes, and the XGB model performed better than the LR model.

Conclusion:

Machine learning algorithms using prehospital information showed satisfactory performance in diagnosing stroke and predicting the need for intra-arterial thrombolysis.