Preterm infant general movements assessment via representation learning

General movements assessment (GMA) is essential for abnormal movement screening and timely intervention of neuromotor diseases for preterm infants. Conventional preterm infant GMA requires long-term infant movement observation by medical experts with domain-specific knowledge, which is labor-intensive. To bridge the gap between time-consuming observation-based conventional GMA and cost-effective video-based automatic GMA for preterm infants, we contribute a smartphone video dataset dedicated to full-body movement of 87 preterm infants and propose a video-based preterm infant abnormal movement detection system that can be generalized to preterm infant GMA beyond our dataset. Our system first extracts the trajectories of the body key-points with HRNet to learn the representations that are most relevant to the movement while avoiding the irrelevant appearance-related features, and then returns a normal/abnormal movement binary prediction with CTR-GCN for automatic preterm infant GMA. Unlike most existing works that adopts the mixture of full-term and preterm infants or the second-stage preterm infants for GMA, our dataset purely focuses on the first-stage preterm infants for in-time prediction. Experimental results show that our system outperforms the competing GMA schemes and achieves a 95.54% accuracy rate in terms of the detection of intra-subject potential neuromotor diseases.