Movement Pattern Analysis Based on Point-Line-Plane Hierarchies and Machine Learning for Fall Risk Assessment in Community-Dwelling Older Adults.

This study systematically compared full-body, upper limb, lower limb, and trunk kinematic features in fall risk classification among community-dwelling older adults. Full-body and upper limb features yielded higher accuracy than lower limb and trunk, underscoring the key role of arm movements in balance control. Feature importance analysis selected 12.5% of key variables, including wrist displacement and upper limb velocity, boosting model accuracy from 57% to 78%, demonstrating strong clinical potential. Different machine-learning models showed complementary strengths: XGBoost excelled with nonlinear lower limb features, while random forest better integrated heterogeneous full-body data. By contrast, traditional linear kinematic-based models achieved a maximum accuracy of only 0.61, reflecting their limited ability to capture the nonlinear, multiscale, and sensory integration aspects of postural control. Integrating multi-regional movement coordination enhanced prediction, highlighting the multidimensional nature of balance regulation and supporting the development of efficient clinical fall risk screening tools.