February 09, 2023
1 min read
Raychev RI, et al. Abstract WMP120: Development of smartphone enabled machine learning algorithms for autonomous stroke detection. Presented at: International Stroke Conference; Dallas; Feb. 8-10, 2023.
Raychev reports receiving a research grant for the current work, support from Modest and the Society of Vascular and Interventional Neurology, and ownership interest in Modest and Spartan Micro. Please see the study for all other authors’ relevant financial disclosures.
A smartphone enabled machine-learning algorithm may be as effective as a neurologist in identifying signs and symptoms of acute stroke, according to preliminary research presented at International Stroke Conference.
“Many stroke patients don’t make it to the hospital in time for clot-busting treatment, which is one reason why it is vital to recognize stroke symptoms and call 9-1-1 right away,” Radoslav I. Raychev, MD, FAHA, clinical professor of neurology at the University of California, Los Angeles, stated in a related press release.
Raychev and colleagues developed FAST.AI, a smartphone application designed to recognize stroke using machine learning algorithms, which identified typical symptoms such as facial asymmetry, upper limb weakness and alterations in speech.
Researchers analyzed data from 269 individuals diagnosed with acute stroke (median age, 71 years; 41% women) who were admitted to four major metropolitan stroke centers in Eastern Europe between July 2021 and July 2022. Data capture of speech patterns and facial expressions occurred through video recording, with arm data collected by device sensors.
Elements of the algorithm include 68 facial landmark points to measure asymmetry, a grasp agnostic classifier to detect arm weakness and a frequency-analysis component to detect abnormal or slurred speech. Researchers conducted all tests within 72 hours of admission and compared each machine-learning output with neurologists’ clinical impression.
According to results, analyses of 18,311 facial images demonstrated a significant degree of sensitivity (99.42%), specificity (93.67%) and accuracy (97.11%) in detecting facial asymmetry, while results of 43 motion trajectories detected arm weakness with 71.42% sensitivity, 72.41% specificity and 72.09% accuracy.
Researchers also reported that preliminary analysis of speech-alteration algorithms confirmed adequate features to detect abnormalities.
“Early results confirm the app reliably identified acute stroke symptoms as accurately as a neurologist, and they will help to improve the app’s accuracy in detecting signs and symptoms of stroke,” Raychev said in the release.