Scaling Dynamic Mode Decomposition for Real Time Analysis of Infant Movements

Abstract

The analysis of characteristic motions in infants plays a pivotal role in quantifying developmental progress and clinical risk for neurodevelopmental and musculoskeletal abnormalities. Traditional methods often rely on resource intensive manual motion assessments carried out by clinicians while computer assisted approaches frequently utilize computationally expensive simulations or black-box classification models. These approaches struggle to efficiently to both capture and differentiate the highly correlated dynamics of infant motion, limiting their ability to deliver actionable insights in a clinically viable decision time frame. In response to these challenges, we introduce the use of Dynamic Mode Decomposition (DMD) as a transformative approach for decomposing complex infant motion into interpretable, independent components that are linearly additive in nature. DMD not only enables extraction of large scale clinically meaningful patterns but also can integrate with existing computer assisted interventions with regard to standardized motion features. We assess an optimized DMD formulation on 275,000 frames of infant motion in clinical settings that have undergone manual motion assessment by clinicians. Our experimental results show that using DMD modes as predictive components not only result in equal or superior accuracy in predicting abnormal clinical motion assessments compared to traditional manual or computer assisted methods but serve as highly data rich features themselves that can be used as a novel basis for personalized clinical analysis and uncertainty quantification at scale.