Infant neurodevelopmental impairment detection via large-scale motion screening with generative learning

Neurodevelopmental impairments like cerebral palsy are prevalent infant developmental problems in clinical infant healthcare. Early detection of these disorders enables timely intervention and better rehabilitation outcomes. Automated, large-scale screening tools remain scarce, and artificial intelligence offers a promising solution by identifying abnormal infant motion patterns. However, developing a data-driven AI system is challenging due to the scarcity of expert-annotated motion records. To address the challenge, this paper proposed a generative pattern learning method, which achieves sensitive detection of abnormal actions through the training paradigm of temporal auto-regression and reconstruction. We trained and validated the AI model DeepAD with more than 1,300 infant motion videos over a period of 5 years (2019-2024) from three children’s hospitals in China, which contain about 50 infants diagnosed with cerebral palsy. Extensive experiments demonstrate the superior performance of DeepAD over baseline methods and equivalent level to clinical experts, indicating its potential for smart screening of neurodevelopmental disorders.