Fall detection algorithm and wearable inertial sensor location optimization

Abstract

Falls are one of the major health and safety concerns in the elderly population, with around 30% yearly fall rate for those over 65 years old. As the elderly population is growing rapidly, falls are not only affecting the elders’ lives but also imposing a burden on the healthcare system. Over the past decade, many researchers developed different fall detection algorithms using wearable inertial sensors. To find the best fall detection algorithms, previous researchers compared multiple algorithms using different datasets from different sensor locations. However, most of datasets have limited types of motions from small number of subjects. Thus, their algorithm accuracy may have been overestimated and generalization is questionable. In addition, few comparison studies have covered all three major types of algorithms (threshold-based, conventional machine learning, deep learning) for fall detection, especially with both accuracy and practicality measures. Furthermore, most of them used either threshold-based or machine learning algorithms to compare the algorithms, so that the optimal sensor location may be found differently depends on the algorithms applied. This study aims to comprehensively compare three different types of algorithms for fall detection and optimize the sensor location. First, a large-scale motion dataset from 32 young subjects wearing inertial sensors in waist, chest, and upper leg location was built with 21 types of activities of daily life (ADLs) and 15 types of falls. Based on the common dataset from waist location, which was used as a reference location, threshold-based algorithm, conventional machine learning algorithm, and deep learning algorithm were applied to compare the fall detection performance. For each type of algorithm, four thresholds-based, Support vector machine (SVM), and convolutional neural network (CNN) were chosen and compared thoroughly in terms of both accuracy and practicality measures. After the comparison, the best-performed algorithm was applied to the other two sensor locations, and the performances in three sensor locations were compared. The first study revealed that CNN outperformed the other two algorithms with 100% accuracy and timely detection within 119.7ms after fall impact. The second study showed that CNN on the waist performed best, followed by CNN on the chest, and CNN on the upper leg as last. Our research findings suggest that CNN has great potential in fall detection, and to develop a fall detection application, waist is a promising sensor location to consider. This study could provide a good reference for the researchers and industries in the field of healthy aging when designing fall detection and alarm systems.