Human vs. Machine Learning Based Detection of Facial Weakness Using Video Analysis.

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Imaging and Computer Vision
• Neurology and Stroke Diagnostics

Background:

• Current emergency medical services (EMS) stroke screening tools have variable accuracy and reliability.
• An automated tool could enhance prehospital stroke diagnosis and patient outcomes.
• Facial weakness is a common stroke symptom that may be detectable via video analysis.

Purpose of the Study:

• To test the hypothesis that a machine learning algorithm can detect facial weakness using video analysis.
• To compare the algorithm's performance against trained paramedics in identifying facial weakness and its laterality.

Main Methods:

• Curated 93 videos of individuals with unilateral facial weakness and 96 videos of individuals with normal smiles from public sources.
• Had neurologists and paramedics independently assess videos for facial weakness and laterality.
• Trained a computer vision algorithm using a 5-fold cross-validation scheme on the curated video dataset.

Main Results:

• Paramedics achieved a mean accuracy of 92.6%, sensitivity of 87.8%, and specificity of 99.3% in video analysis.
• The machine learning algorithm demonstrated an accuracy of 88.9%, sensitivity of 90.3%, and specificity of 87.5%.
• Algorithm's sensitivity for detecting weakness was comparable to paramedics, though specificity was lower.

Conclusions:

• Machine learning and computer vision show promise for detecting unilateral facial weakness in videos.
• The algorithm's performance is comparable to trained paramedics in terms of accuracy and sensitivity.
• Further research and external validation are needed for prospective patient encounters and augmented detection.