Quantifying Arm and Leg Movements in 3-Month-Old Infants Using Pose Estimation: Proof of Concept

Abstract

Background: Pose estimation (PE) has the promise to measure pediatric movement from a video recording. The purpose of this study was to quantify the accuracy of a PE model to detect arm and leg movements in 3-month-old infants with and without (TD, for typical development) complex congenital heart disease (CCHD). Methods: Data from 12 3-month-old infants (N = 6 TD and N = 6 CCHD) were used to assess MediaPipe’s full-body model. Positive predictive value (PPV) and sensitivity assessed the model’s accuracy with behavioral coding. Results: Overall, 499 leg and arm movements were identified, and the model had a PPV of 85% and a sensitivity of 94%. The model’s PPV in TD was 84% and the sensitivity was 93%. The model’s PPV in CCHD was 87% and the sensitivity was 98%. Movements per hour ranged from 399 to 4211 for legs and 236 to 3767 for arms for all participants, similar ranges to the literature on wearables. No group differences were detected. Conclusions: There is a strong promise for PE and models to describe infant movements with accessible and affordable resources—like a cell phone and curated video repositories. These models can be used to further improve developmental assessments of limb function, movement, and changes over time.