AI Mobile Application Versus HCP for Bodyweight Squats

Artificial Intelligence (AI) Mobile Application Versus Health Care Provider (HCP) for Bodyweight Squats: A Randomized, Blinded, Controlled Clinical Trial

To assess if an artificial intelligence (AI) mobile application can identify and improve bodyweight squat form in adult participants when compared to a Physical Therapist (PT).

Artificial intelligence (AI) is changing the way people can address their health needs. One such way related to physical exercise is AI-enabled exercise mobile application (digital coach), which uses motion tracking technology to monitor and provide real-time audio feedback on a person's exercise form. However, this AI technology has yet to be independently tested against an in-person evaluator (human coach) for its ability to improve exercise form. This study is a blinded randomized controlled trial comparing the ability of the digital coach (n=15) and a Physical Therapist (PT) human coach (n=15) to improve bodyweight squat form in 30 able-bodied volunteers age 20 - 35. Each volunteer performs 10 unassisted control squats, then 10 squats with assistive vocal feedback from either coach after each repetition, and finally 10 more unassisted test squats, all squats video-recorded. Three independent video evaluators count the number of correct squat repetitions completed by volunteers before and after intervention by the different coaches. This project is important to validate the digital coach compared to a PT human coach in a small population using a bodyweight squat for its wide applicability to daily movement patterns.

To determine baseline ability and serve as their own control, participants in both groups performed 10 bodyweight squat "control" repetitions without feedback followed by one minute of rest. Those in the AI group then performed 10 more "practice" repetitions with real-time audiovisual feedback from the app followed by one minute of rest. The AI's design provided one piece of feedback, if necessary, with a vocal statement and on-screen video per repetition (e.g. when a participant performed a squat repetition with their neck flexed downward, AI suggested keeping their head up with on-screen instruction). Participants in both groups then performed 10 "test" repetitions without feedback followed by one minute of rest.

To determine baseline ability and serve as their own control, participants in both groups performed 10 bodyweight squat "control" repetitions without feedback followed by one minute of rest. Those in the PT group (n=15) also performed 10 "practice" repetitions with one piece of feedback per repetition, if necessary, from the PT followed by one minute of rest. Participants in both groups then performed 10 "test" repetitions without feedback followed by one minute of rest.

Post-intervention improvement in squats will be determined by the number of correct squats in the third set as compared to the first set of squats.

AI identification of correct and incorrect squats will be determined by the number of squats that are identified correctly by AI as compared with independent evaluators.

Inclusion Criteria: Columbia University affiliate Aged 20 to 35 years Able to perform moderate bodyweight exercise for 10 minutes Exclusion Criteria: Unable to provide consent

Luna A, Casertano L, Timmerberg J, O'Neil M, Machowsky J, Leu CS, Lin J, Fang Z, Douglas W, Agrawal S. Artificial intelligence application versus physical therapist for squat evaluation: a randomized controlled trial. Sci Rep. 2021 Sep 13;11(1):18109. doi: 10.1038/s41598-021-97343-y.