Abstract WP246: Automated Detection of Facial Weakness Using Machine Learning
Autor: | Iris Lin, Mark McDonald, Gustavo K. Rohde, William Alex Dalrymple, Bradford B. Worrall, Omar Uribe, Yan Zhuang, Daniel F Arteaga, Andrew M. Southerland |
---|---|
Rok vydání: | 2018 |
Předmět: | |
Zdroj: | Stroke. 49 |
ISSN: | 1524-4628 0039-2499 |
DOI: | 10.1161/str.49.suppl_1.wp246 |
Popis: | Background: Early recognition and treatment of stroke improves outcomes. Pre-hospital screening tools offer promise for early detection of acute stroke but have inconsistent performance. An automated screening tool could minimize inter-operator variability and operator error. We hypothesize that machine learning algorithms can assist in the detection of pathologic facial weakness using computational visual analysis. Methods: Two senior neurology residents scored images (n = 333) showing normal smile and facial weakness on a 5-point scale. Only images rated concordantly by both raters as likely normal or likely abnormal were included for analysis. Facial landmarks were extracted using an open-source face-detection and landmark localization algorithm. A penalized linear discriminant analysis method was used to process the data with a k-nearest neighbor algorithm as the classifier. We used a 5-fold cross-validation scheme and calculated accuracy, sensitivity, and specificity to evaluate performance. Results: Of the 199 images analyzed, 18 images were excluded due to limitations in facial landmark extraction. Of the remaining 181 images, 87 were rated as likely normal and 94 were rated as likely having facial weakness. The algorithm performed with 93.8% accuracy (95% CI [93.4-94.2]), 95.8% sensitivity [95.4-96.2] and 93.8% specificity [93.4-94.2]. Discussion: In this pilot study, we demonstrate that machine learning algorithms can accurately identify facial weakness in static images. These results support further evaluation of machine learning algorithms in the early detection of acute stroke symptoms. Future research will examine video analysis algorithms in detecting facial weakness and other signs in acute stroke patients. |
Databáze: | OpenAIRE |
Externí odkaz: |