This study aimed for two purposes: (1) to develop and validate AI algorithms for classifying movements in full-term and preterm infants when video recorded in a clinical laboratory and at home via a mobile App during early childhood, and (2) to organize the movements classified by the AI algorithms into age-based sets and examine their concurrent validity with physiotherapists' assessment results.
The study included 85 full-term infants and 82 preterm infants for prospective examination of the Alberta Infant Motor Assessment (AIMS) in a laboratory from 4 to 18 months of age and invited the parents to upload videos at home via the App “Baby Go” for later movement classification. The AIMS contains 58 movements (21 in prone, 9 in supine, 12 in sitting, and 16 in standing). All video records were annotated by three physiotherapists and used as the gold standard to examine the accuracy of the AI model. The movements recognized in the AI model were dispatched into age-based sets to test concurrent validity using the AIMS assessment results as the criterion measure.
The infants contributed to 510 AIMS assessment sessions and 2,515 video records in the laboratory, further sliced into 36,805 data samples; 82 parents uploaded 1,376 video records via the APP from home, further sliced into 2,599 data samples. Validation of the action recognition model for classifying 38 movements in full-term and preterm infants combined showed an overall accuracy of 0.91, precision of 0.92, recall of 0.90, and F1 score of 0.91 for the laboratory videos, and accuracy of 0.84, precision of 0.84, recall of 0.77, and F1 score of 0.78 for the home videos. The AI algorithms showed similarly high accuracy in classifying these movements in full-term and preterm infants, respectively, in the laboratory videos, and so did the home videos. Furthermore, the 38 movements were dispatched into age-based sets with two to five movements per age that showed high concurrent validity with the AIMS results in all infants across ages (agreement = 0.99, sensitivity = 0.92, specificity = 0.99, positive predictive value = 0.86, and negative predictive value =1.00).
The AI model accurately classifies 38 movements in full-term and preterm infants performed in the laboratory and at home. The age-based sets also correlate highly with the physiotherapists’ assessment results.
These findings provide the basis for future validation of the combined AI algorithm and APP for remote infant motor assessment.
Mobile Application
Motor Development
