A pose estimation for motion tracking of infants cerebral palsy

Abstract

The General Movements Analysis (GMA) has demonstrated noteworthy promise in the early detection of infantile Cerebral Palsy (CP). However, it is subjective and requires highly trained clinicians, making it costly and time-consuming. Automation of GMA could potentially enhance accessibility and further our comprehension of infants’ fullbody movements. This paper investigates the feasibility of using 2D and 3D pose estimation strategies to observe and scrutinize the infant’s comprehensive body movement attributes to improve our perspective to consider joint movement and positions over time as an alternative to GMA for early CP prediction. The study includes comprehensive movement analysis from video recordings for accurate and efcient analysis of infant movement by computing various metrics such as angle orientations at diferent predicted joint locations, postural information, postural variability, movement velocity, movement variability, and left–right movement coordination. Along with antigravity movements are assessed and tracked as indicators of CP. We employed a variety Machine Learning (ML) algorithms for CP classifcation based on a series of robust features that have been developed to enhance the interpretability of the model. The proposed approach is assessed through experimentation using the MINI-RGBD and RVI-38 datasets with a classifcation accuracy of 92% and 97.37% respectively. These results substantiate the efcacy of employing pose estimation techniques for the precocious prediction of infantile CP, highlighting the importance of monitoring changes in joint angles over time for accurate diagnosis and treatment planning.