606 - Computer Vision-Based Detection of Movement Patterns Predictive of Neuromotor Conditions in Newborns

Publication Number: 3587.606

Marissa Koscielski, University of Cincinnati College of Medicine, cincinnati, OH, United States; Patrick Tinsley, AngelEye Health, Hoboken, NJ, United States; Adrian Rodriguez, AngelEye Health, Arcadia, CA, United States; Kathleen Adderley, Nationwide Children's Hospital, Columbus, OH, United States

Background: Neuromotor conditions arising in childhood represent a leading cause of long-term motor impairment, disability, and healthcare burden. Although early detection and intervention studies have demonstrated improved outcomes, traditional neuromotor assessments pose significant hurdles for universal screening. The advent of sensor-based analysis movement pattern categorization may be a means to support clinicians in detecting movement patterns predictive of neurologic function in infants.

Objective: The purpose of this study is to analyze the efficacy of machine learning models for differentiating atypical movement patterns from typical movement patterns, as defined by neuromotor diagnoses at two years of age.

Design/Methods: In this IRB-approved study, prospective video data was collected from infants aged 0–6 months corrected age. At two year follow-up, a retrospective chart review pulled all diagnoses and neurologic and motor findings to categorize the infant as typical or atypical. Pose estimate data for all videos was extracted and kinematic features were constructed. Temporal deep learning models were then trained to classify samples and the associated movement patterns relative to outcome groups. Model performance was assessed in a cross-validation setup with AUC, sensitivity, and specificity as the reported metrics.

Results: Of the 1185 infants enrolled in the study with complete data, 361 were classified as atypical. Diagnoses included autism (79), cerebral palsy (47), spinal muscular atrophy (26), Down syndrome (22), and dystrophinopathy (5). The outstanding samples were labeled with multiple diagnoses (15) or as “Other” (167). Pilot models trained on the most informative kinematic features achieved a median AUC of 0.72. The most predictive features included features that measured the distance between pairs of joints and features that captured raw positional data. Subgroup analyses focused on distinguishing cerebral palsy (CP) samples from typical samples produced models with median AUC of 0.82.

Conclusion(s): In this study, computer vision techniques were able to distinguish typical from atypical movement patterns in early infancy that are predictive of neuromotor diagnoses at two years. With further research, this approach may have potential to support clinicians with information for earlier screening and detection of atypical movement patterns, ultimately improving long-term outcomes. Additional analyses are underway to further determine the most predictive features of specific neuromotor diagnoses.