Home Alone Training: Workout Effectively with Deep Learning Based Mobile Assistant (Preprint)

BACKGROUND Insufficient physical activity due to social distancing and suppressed outdoor activities increase people’s vulnerability against various diseases such as sarcopenia and severe COVID-19 illness. While bodyweight exercises like squats are effective in increasing the amount of activity at home while refraining from outdoor activities, incorrect postures can activate wrong muscles, potentially leading to ineffective sessions or even injuries, and avoiding incorrect postures is not an easy task for novices without the help of an expert. Although remote coaching or computer-assisted posture correction systems have been introduced, most existing solutions are costly or inefficient. OBJECTIVE We aim to use deep neural networks to design and develop a personal workout assistant capable of giving feedback on squat postures using only mobile devices such as smartphones and tablets. We also evaluate the effectiveness of our mobile system by comparing it with simply following exercise videos, which is one of the most popular training methods at home. METHODS In the first part, we created a dataset with over 20,000 squat videos, annotated by experts, and trained a deep learning model using a combination of pose estimation and video classification to analyze the workout postures. In the second part, we developed a mobile workout assistant application and showed how our application helps people improve squat performance via an interventional study for two weeks in which the experimental group trained using our assistant application while the control group only watched and followed workout videos. RESULTS The participants in the group that used our application significantly improved their squat performance evaluated by the application after two weeks (P=.001) while the control group without the application had no significant change in performance (P=.13). There were significant differences in the left (P=.02) and right (P=.02) knee’s joint angles in the experimental group between the before and after, while no significant effect was found for the control group in the left (P=.68) and right (P=.61) knee’s joint angle. CONCLUSIONS While both groups managed to increase their muscle strength and endurance the participants in the experimental group who trained with our mobile assistant application experienced faster improvement and learned more nuanced details of the full squat exercise without additional guidance. CLINICALTRIAL This study does not include a clinical trial.