Development and evaluation of an elderly fall detection system based on a wearable device located at wrist

Abstract

Falls in the elderly age are a world health problem. Every year, about 30% of people aged 65 or older become victims of fall events. The consequences of a fall may be physiological (e.g. bone fractures, muscular injuries) and psychological, including the loss of self-confidence by fear of falling, which leads to new falls. A solution to this problem is related to preventive actions (e.g. adapting furniture) allied to fall detection systems, which can alert family members and emergency medical services. Since the response time for help is related to the fall's consequences and severity, such systems must offer high accuracy and real-time fall detection. Although there are many fall detection solutions in literature (most part of them related to wearable devices), few of them are related to wrist-worn devices, mainly because of the existing challenges for this configuration. Considering the wrist as a comfortable, discrete and acceptable place for an elderly wearable device (less associated to the stigma of using a medical device), this work proposes the development and evaluation of a fall detection solution based on this configuration. For this, different sensors (accelerometer, gyroscope and magnetometer) were combined to different algorithms, based on threshold and machine learning methods, in order to define the best signals and approach for an elderly fall detection. These methods considered acceleration, velocity and displacement information, relating them with wrist spatial orientation, allowing the calculation of the vertical components of each movement. For the algorithms' training and evaluation, two different protocols were employed: one involving 2 volunteers (both males, ages of 27 and 31) performing a total of 80 fall and 80 non-fall events simulation, and the other involving 22 volunteers (14/8 males/females, ages mean: 25.2 ± 4.7) performing a total of 396 fall and 396 non-fall events simulation. An exhaustive evaluation of different signals and configuration parameters was performed for each method. The best threshold-based algorithm employed the vertical acceleration and total velocity signals, achieving 95.8% and 86.5% of sensitivity and specificity, respectively. On the other hand, the best machine learning algorithm was based on the K-Nearest Neighbors method employing the vertical acceleration, velocity and displacement information combined with spatial orientation angles: 100% of sensitivity and 97.9% of specificity. The obtained results allow to emphasize the relevance of machine learning algorithms for wrist-worn fall detection systems instead of traditional threshold-based algorithms. These results offer great contributions for the research of similar wearable fall detectors, suggesting the best approach for new developments.