Impact of Gait Parameters and Their Variability on Fall Risk Assessment Accuracy Using Wearable Sensor

Wearable sensors are increasingly utilized in fall risk assessments, providing precise stride-to-stride spatiotemporal gait parameters that are correlated with a heightened risk of falls. However, the impact of these gait parameters and their variability on the overall accuracy of fall risk prediction models remains an open question. This study introduced three fundamental machine learning models-logistic regression, support vector machines (SVM), and an artificial neural network-to predict fall risk among 163 frail older adults. Gait parameters and their variability were collected from a foot-mounted inertial measurement unit (IMU) and computed based on walking test durations ranging from 1 to 15 minutes, instead of using stride numbers, which are impractical in real clinical settings. Leave-one-out cross-validation was employed to evaluate the models' performance, revealing that optimal walking test durations ranged from 6 to 10 minutes. The artificial neural network demonstrated the highest accuracy, achieving a score of 0.96 during an 8-minute test. These findings provide critical insights for designing experimental protocols in fall risk assessments using wearable technology.