Abstract
Neurodevelopmental impairments (NDIs) are significant long-term complications in preterm infants. While early recognition of infants at high risk for NDIs is essential for enabling timely intervention, it remains a challenging endeavor. Autonomous screening methods utilizing sensors represent a promising avenue for addressing this challenge, yet they remain an open area of research. This study presents a novel frequency modulated continuous wave (FMCW) radar-based machine learning (ML) system designed for early screening to predict and identify infants at high risk for poor neurodevelopmental outcomes. The proposed method constructs 3-dimensional (3D) range-angle-time data cube of infant movements using the radar sensor. Based on this 3D radar data cube, asymmetric movements are identified by analyzing the ratio of left and right movements, while an ML model detects abnormal movements (cramped-synchronized general movements, CSGMs). Using the frequency of asymmetric movements and CSGMs, we propose a new index, termed "neuroriskability (NRA)", which ultimately determines the overall risk of NDIs. The NRA scores generated from the radar data were compared with clinically evaluated neurodevelopmental outcomes through experiments conducted with both hospitalized and outpatient infants to validate their clinical utility. The proposed method successfully predicted infants with poor neurodevelopmental outcomes, thereby demonstrating its feasibility in clinical practice.