Artificial Intelligence for Infant Motor Screening: Development and Validation

Background and Purpose. Although the number of children with developmental disorders reported for early intervention in Taiwan increases in the recent decade, the prevalence estimate of children with developmental disorders is lower than the global data particularly among those aged under 2 years or in remote areas. Artificial Intelligence (AI), based on machine learning of big data, has been successfully used for medical image classification and prediction in certain diseases; however, its application in child developmental screening is rare. The purpose of this three-year study is therefore three-fold: (1) Model Development- to apply pose estimation model and tracking recognition model on the movements of a large sample of term and preterm infants under a motor assessment in the laboratory to examine the accuracy of the AI algorithms in identifying individual movements using physical therapists' results as gold standards; (2) Model Validation- to examine the performance of the AI algorithms on the same term and preterm infants' movements when video recorded by the parents at home between the laboratory assessment ages using physical therapists' results as gold standards; and (3) Concurrent and Predictive Validity of AI Movement Sets- to select the identifiable movement classes into AI movement sets for individual ages to examine their concurrent validity with physical therapists' results and predictive validity on developmental outcomes at 18 months of age in these infants. Method. A total of 125 term and preterm infants will be recruited from National Taiwan University Children's Hospital and will be randomly split into the training (N=101), tuning (N=12), and testing sets (N=12) with 8:1:1 ratio for Model Development. All infants will be prospectively administered the Alberta Infant Motor Assessment in prone, supine, sitting and standing positions at 4, 6, 8, 10, 12 and 14 months of age (corrected for prematurity) in the laboratory with movements recorded by 5 cameras. For Model Validation, the same 125 infants will be video recorded their movements by the parents using cell phones at home at 5, 7, 9, 11 and 13 months of age from at least 2 camera views, with the movement records uploaded to a prototype of Mobile APP "Baby Go." The data processing of movement video records will include: selection of movement records, establishment of a pose estimation model, and establishment of an action recognition model. The accuracy of the AI model in identifying infants' individual movements will be examined using physical therapist's results as gold standards. The movements identifiable through machine learning will be selected to establish AI movement sets for each age. Concurrent and Predictive Validity of the AI movement sets will be respectively examined using physical therapist's results and developmental outcomes at 18 months of age as the criteria (age of walking attainment and the Peabody Developmental Motor Scale- 2nd edition). Significance. The results will help establish the best and appropriate AI model for infant motor screening in Taiwan. The established AI model may be incorporated into clinical procedure to assist pediatricians and physical therapists in planning for further diagnostic assessment.