Using DeepLabCut to Recognize Early Motor Development Patterns Associated with Neurodevelopmental Disorders

Early identification of Neurodevelopmental Disorders (NDD) allows for faster intervention, which in turn improves clinical outcomes and reduces the individual and societal costs associated with the diagnosis. The aims of the study were to 1) investigate the use of the DeepLabCut (DLC) toolbox to automatically analyze the motor patterns of infants at Low Risk (LR) and High Risk (HR) for Autism Spectrum Disorder (ASD); and 2) define the critical time window in which atypical motor patterns discriminate between typically developing infants and those diagnosed with ASD or NDD. The DLC toolbox was used to train a model capable of tracking the movements of both LR and HR infants longitudinally at the ages of 10 days, 6 weeks, 12 weeks, 18 weeks, and 24 weeks. 226 videos of 87 infants (45 females), collected within the Italian Network for Early Detection of Autism Spectrum Disorder (NIDA), were analyzed. Using the Percentage of Correct Key-points (PCKh) accuracy metric, the DLC's tracking performance was verified by comparing the obtained 2D hands and feet coordinates with those extracted by the Movidea software. Furthermore, motor features were computed and fed to three classifiers: Fine Tree, RUSBoosted Trees, and Narrow Neural Network to investigate their usefulness in terms of early NDD prediction. Satisfactory PCKh results were obtained for both hands and feet (left foot: 96.6%, right foot: 96.2 %, left hand: 80.9%, right hand: 82.8%). The best classification results were obtained with the RUSBoosted classifier at the ages of 10 days and 6 weeks. The 5-fold cross-validation accuracy was 81.4%, with a true negative rate of 80.0% and true positive rate 87.5%. Our data confirm the usefulness of DLC as a low-cost approach to track infant movements during the writhing period. Early motor behavior at the ages of 10 days and 6 weeks carries valuable information that has the potential to be suitable in predicting the diagnosis of NDD.