An Automated Framework for Pose-Sequence Analysis of Spontaneous Movements in Preterm Infants

Goal: Spontaneous movement analysis is crucial for early neuromotor assessment, yet objective, unobtrusive, and scalable tools remain limited. This study investigates an automated video-based framework for characterizing spontaneous movements in preterm infants. Methods: Videos of 142 infants at term-equivalent age were analyzed using markerless two-dimensional pose estimation. Pose configurations were clustered into a pose dictionary, and recordings were represented as temporal sequences of key poses from which movement parameters were derived. Infants with typical development (TD) were compared with those with cerebral palsy (CP) and developmental delay (DD) stratified into low and high motor severity. Results: Compared with TD, CP and DD-high showed significantly lower pose transition (median: 0.38 and 0.42 vs 0.50 transitions/s) and accumulated fewer unique poses over time, while CP also showed lower pose variability (median: 0.79 vs 0.87). Conclusions: The framework characterizes pronounced neuromotor impairments and may provide an objective approach for early motor characterization.