Machine Learning to Identify a New Digital Biomarker to Monitor Everyday Upper Limb Use in Children with Unilateral Cerebral Palsy

Abstract

Children with Unilateral Cerebral Palsy (UCP) often experience reduced spontaneous use of the non-dominant upper limb in daily life. Traditional clinical assessments, such as the Assisting Hand Assessment (AHA), provide valuable but clinical-environment-related and thus episodic measures of functional performance in structured settings. There remains a critical need for tools that enable continuous, ecologically valid, and objective monitoring of upper limb activity in real-world environments. In this study, we introduce the Daily AHA Biomarker (DAB) , a novel digital biomarker derived from wearable sensor data, designed to estimate AHA scores based on spontaneous motor behavior. The DAB is intended to be used in clinical practice to monitor the movement of the upper extremities of subjects with UCP in their naturalistic environments. Using bilateral wrist-worn accelerometers, we collected multiday time-series data from 80 children (54 with UCP and 26 with Typical Development). Our Machine Learning pipeline combines time-series classification and regression to predict AHA scores from unstructured, daily living-recorded data. The final DAB indicator showed high predictive accuracy ( $R^2 = 0.709$ ) and a strong correlation with the clinical AHA score and the Manual Ability Classification System (MACS) level.