Machine Learning-based Detection of Compensatory Balance Responses and Environmental Fall Risks Using Wearable Sensors
MetadataShow full item record
Falls are the leading cause of fatal and non-fatal injuries among seniors worldwide, with serious and costly consequences. Compensatory balance responses (CBRs) are reactions to recover stability following a loss of balance, potentially resulting in a fall if sufficient recovery mechanisms are not activated. While performance of CBRs are demonstrated risk factors for falls in seniors, the frequency, type, and underlying cause of these incidents occurring in everyday life have not been well investigated. This study was spawned from the lack of research on development of fall risk assessment methods that can be used for continuous and long-term mobility monitoring of the geri- atric population, during activities of daily living, and in their dwellings. Wearable sensor systems (WSS) offer a promising approach for continuous real-time detection of gait and balance behavior to assess the risk of falling during activities of daily living. To detect CBRs, we record movement signals (e.g. acceleration) and activity patterns of four muscles involving in maintaining balance using wearable inertial measurement units (IMUs) and surface electromyography (sEMG) sensors. To develop more robust detection methods, we investigate machine learning approaches (e.g., support vector machines, neural networks) and successfully detect lateral CBRs, during normal gait with accuracies of 92.4% and 98.1% using sEMG and IMU signals, respectively. Moreover, to detect environmental fall-related hazards that are associated with CBRs, and affect balance control behavior of seniors, we employ an egocentric mobile vision system mounted on participants chest. Two algorithms (e.g. Gabor Barcodes and Convolutional Neural Networks) are developed. Our vision-based method detects 17 different classes of environmental risk factors (e.g., stairs, ramps, curbs) with 88.5% accuracy. To the best of the authors knowledge, this study is the first to develop and evaluate an automated vision-based method for fall hazard detection.