Abstract Details

Title On Evaluating The Potential Of Machine Learning For Effective Classification Of Ischemic And Hemorrhagic Stroke In Resource Limited Settings

Topic Cerebrovascular Disease and Interventional Neurology

Presentation(s) P1 - Poster Session 1 (9:00 AM-5:00 PM)

Poster/Presentation
Number 114

Objective To evaluate the capabilities of Machine Learning for effective differentiation between Ischemic and Hemorrhagic Stroke.

Background Stroke is the second leading contributor to Disability Adjusted Life Years and most of the stroke burden is over developing nations. A major factor for reducing stroke burden is timely identifying stroke and the type of stroke as either ischemic(IS) or hemorrhagic(HS), as early identification of IS should shorten the workflow of acute stroke management. Where neuroimaging is unavailable, several clinical scores were developed with limited utility in clinical practice to identify stroke type. Several studies show that these existing scores fail to perform for different demographics.

Design/Methods We collect retrospective dataset of 422 subjects(IS=324, HS=98 diagnosed by CT) and 149 clinical attributes to evaluate the performance of three classifiers namely BalancedRandomForest, AdaBoost, GradientBoosting, and their ensemble. For class imbalance, we recommend Weighted Accuracy(WA) as performance measure.

Results

Gradient Boosting achieves highest accuracy (86.97%) and sensitivity (96.29%). Moreover, ensembling the three classifiers further improved performance on accuracy without compromising much on WA. BalancedRandomForest is a more effective classifier as it achieves highest WA (82.33%) and has comparable sensitivity (80.02%) and specificity (81.63%). This implies that it is equally effective for both IS and HS diagnosis.

Classifier	Sensitivity (IS)	Specificity (IS)	Accuracy	WA
BalancedRandomForest	83.02%	81.63%	82.70%	82.33%
AdaBoost	94.44%	55.10%	85.31%	74.77%
GradientBoosting	96.29%	56.12%	86.97%	76.21%
Ensemble	95.37%	65.30%	88.39%	80.34%

Conclusions We demonstrate in this pilot dataset with demographic, clinical features and baseline investigations that ML models may out-perform the existing clinical scores to differentiate IS and HS. Analysis of gait and face using Computer Vision may further enhance the performance. Our classifiers out-perform the existing clinical scores. However, due to unavailability of public databases of previous studies, we could not perform a comparative analysis. Therefore, we recommend authors to make the anonymized datasets publicly available.

Authors/Disclosures
PRESENTER	No disclosure on file
	No disclosure on file
	No disclosure on file
Padma V. Hadakasira, MD (Medwis Healthcare Communications Pvt Ltd)	Dr. Hadakasira has nothing to disclose.
	No disclosure on file
	No disclosure on file
Venugopalan Y. Vishnu, MD (All India Institute of Medical Sciences, New Delhi)	The institution of Dr. Vishnu has received research support from Department of Health Research.
	No disclosure on file

��ɫ��

��ɫ��