Smartphone-Based Fall Risk Assessment Using Postural Feature Analysis and a Dual-Classification Machine Learning Framework

This study proposes a dual-classification approach to fall risk assessment by using machine learning to analyze postural features. Two classification tasks were examined: fall history (Y/N) and real-time fall risk (weak/normal). A range of movement features related to the center of pressure and body keypoints were extracted and analyzed using t-tests, principal component analysis, least absolute shrinkage and selection operator regression, and the machine learning models, random forest and extreme gradient boosting. The results indicate that fall history is associated more with lower-limb features, whereas upper-body dynamics better explain fall risk. The center of pressure-related motion features, relative hand distance, mean distance, and anterior–posterior mean velocity emerged as important variables across the models. Random forest showed superior performance in detecting nonlinear patterns associated with weak/normal classification. We conclude that task-specific model selection and multilevel feature integration are essential for predicting fall risk effectively. Furthermore, camera-based pose estimation from smartphone videos offers a practical and scalable method of real-time fall risk monitoring in home and telehealth settings.