Abstract
<title>Abstract</title> <p>Background: The prevalence of neuromotor conditions, such as cerebral palsy (CP), highlights a critical need for early and scalable detection tools. While expert-based assessments, like the General Movement Assessment (GMA), are effective in the high-risk neonatal population, the screening window and specialized training required of assessors limits widespread use for infants with or without newborn detectable risks. In an effort to address these limitations, this feasibility study explores the efficacy of using computer vision techniques to detect CP-related movements in infants between zero and six months corrected age. Purpose: The purpose of this study is to evaluate the feasibility of using computer vision and time series analysis of infant video data for the task of CP prediction. Methods: In this IRB-approved study, we prospectively collected video data on infants (0–6 months corrected age) and completed chart review at two years of age follow-up to identify children diagnosed with CP. Using an off-the-shelf pose estimation model, we calculated kinematic features such as torso rotation, distances between joints, and angular displacement. Deep learning models were trained on the data to classify infants based on a CP diagnosis at a two-year old follow-up. Model performance was evaluated in a cross-validation experimental set-up with an emphasis on ROC-AUC, sensitivity, and specificity metrics. Results: Out of the 930 infants with complete data, 42 were diagnosed with cerebral palsy at 2 years old. Models that were trained on the most informative kinematic features achieved median ROC-AUC of 0.82, median sensitivity of 0.81, and median specificity of 0.71. The most informative kinematic features were found to be those tracking the distances between two joints. Models trained on manually annotated, salient video segments consistently outperformed those trained on unannotated data. Conclusion: This feasibility study demonstrates the feasibility of applying pose-derived kinematic features and temporal deep learning models for early detection of cerebral palsy in infants. With further research, this approach may have potential to support clinicians with information for earlier screening and detection of aberrant movement patterns, ultimately improving long-term outcomes.</p>