Marker-Less Video Analysis of Infant Movements for Early Identification of Neurodevelopmental Disorders

Background/Objectives: The early identification of neurodevelopmental disorders (NDDs) in infants is crucial for effective intervention and improved long-term outcomes. Recent evidence indicates a correlation between deficits in spontaneous movements in newborns and the likelihood of developing NDDs later in life. This study aims to address this aspect by employing a marker-less Artificial Intelligence (AI) approach for the automatic assessment of infants' movements from single-camera video recordings. Methods: A total of 74 high-risk infants were selected from the Italian Network for Early Detection of Autism Spectrum Disorders (NIDA) database and closely observed at five different time points, ranging from 10 days to 24 weeks of age. Automatic motion tracking was performed using deep learning to capture infants' body landmarks and extract a set of kinematic parameters. Results: Our findings revealed significant differences between infants later diagnosed with NDD and typically developing (TD) infants in three lower limb features at 10 days old: 'Median Velocity', 'Area differing from moving average', and 'Periodicity'. Using a Support Vector Machine (SVM), we achieved an accuracy rate of approximately 85%, a sensitivity of 64%, and a specificity of 100%. We also observed that the disparities in lower limb movements diminished over time points. Furthermore, the tracking accuracy was assessed through a comparative analysis with a validated semi-automatic algorithm (Movidea), obtaining a Pearson correlation (R) of 93.96% (88.61-96.60%) and a root mean square error (RMSE) of 9.52 pixels (7.29-12.37). Conclusions: This research highlights the potential of AI movement analysis for the early detection of NDDs, providing valuable insights into the motor development of infants at risk.