Abstract Details

Title A Machine Learning Approach to Automate Ischemic Stroke Subtyping

Topic Cerebrovascular Disease and Interventional Neurology

Presentation(s) N2 - Neuroscience in the Clinic: Applications of Artificial Intelligence and Machine Learning Tools in Neurology (4:20 PM-4:30 PM)

Poster/Presentation
Number 001

Background Stroke is a leading cause of death worldwide, and optimal management depends on recognizing stroke subtype. However, traditional stroke subtyping is laborious and requires clinical expertise. Since stroke subtypes have not been classified in most large-scale data repositories, automated subtyping would empower efforts to examine subtype determinants and prognosis.

Objective To develop an automated method of applying the Trial of Org 10172 in Acute Stroke Treatment (TOAST) classification system utilizing machine learning on electronic health record (EHR) data (TOAST-AI).

Design/Methods We utilized the EHR data of 1,598 patients with strokes manually subtyped according to TOAST in the Massachusetts General Hospital Ischemic Stroke Registry (2002-2010). In addition to demographics and risk factors, we developed and utilized several novel stroke-specific discriminatory features such as stroke localization, appearance, and vessel stenosis. These features were developed using regular expressions applied on an external dataset of 459,702 radiology reports. We then utilized these EHR features to train a random forest model to discriminate stroke subtypes by TOAST criteria.

Results The analysis sample comprised 360 Definite Cardioembolic (DefCE), 259 Possible Cardioembolic (PosCE), 371 Large Artery Atherosclerosis (LAA), 174 Small Artery Occlusion (SAA), and 278 Undetermined (Undet) strokes. The random forest model showed moderate-to-good classification performance (C-statistic for each subtype: DefCE = 0.93, PosCE = 0.83, LAA = 0.79, SAA = 0.86, and Undet = 0.81). Of the more than fifty features included in the model, those demonstrating greatest influence on performance included atrial fibrillation, patent foramen ovale, lacunar syndrome by neuroradiology report, PR-interval, and severe stenosis of the internal carotid artery.

Conclusions Accurate and efficient stroke subtyping using automated algorithms utilizing EHR data is feasible. Future work is warranted to externally validate our model, apply it in large datasets to facilitate investigation of subtype-specific risk factors, and explore implementation for clinical decision support.

Authors/Disclosures
Ashby Turner, MD (Massachusetts General Hospital) PRESENTER	Dr. Turner has nothing to disclose.
	No disclosure on file
	No disclosure on file
	No disclosure on file
	No disclosure on file
	No disclosure on file
	No disclosure on file
	No disclosure on file
	No disclosure on file
Christopher D. Anderson, MD, PhD, FAAN (Brigham and Women's Hospital)	The institution of Dr. Anderson has received research support from Bayer AG. The institution of Dr. Anderson has received research support from American Heart Association. The institution of Dr. Anderson has received research support from National Institutes of Health. An immediate family member of Dr. Anderson has received publishing royalties from a publication relating to health care.