Comparison of OpenPose and HyperPose artificial intelligence models for analysis of hand-held smartphone videos
Frederick Zhang, Pascale Juneau, Connor McGuirk, Albert Tu, Kevin Cheung, Natalie Baddour, Edward D. Lemaire
Abstract
Movement assessments are invaluable in clinical practice. However, the feasibility of in-person evaluation has been greatly affected due to the COVID-19 pandemic. To overcome this barrier, a virtual assessment system using artificial intelligence (AI) and patient provided videos is needed. AI models for pose inference have produced viable results for identifying a person's joint centers. Identifying AI models for pose inference that provide clinically meaningful results is important for designing a virtual motion assessment tool. This study aims to evaluate the clinical usefulness of two popular pose inference models, OpenPose and HyperPose. Videos recorded by two physicians, who independently performed movements they deemed clinically relevant. Keypoint skeletons were generated and manually inspected frame-by-frame to determine which model produced higher-quality pose inferences. OpenPose produced significantly better scores than HyperPose when comparing within videos (p<; 0.001). Right ankle and right wrist had the poorest performances. Best-practices to be used in the future design of a virtual motion assessment tool are required to improve video "AI-friendliness".