Two Fall-Related and Kinematic Data-Based Approaches for an Instrumented Conventional Cane

Abstract

Stimulating the usability and acceptability of canes within the geriatric community is crucial for fall prevention. Clinical evidences support the use of canes to increase the user's balance and reduce the fall risk. Its nonuse is correlated with a higher number of falls. This fateful event draws tragic consequences such as death or injuries that might affect the victim's psychology, reducing their confidence and mobility. However, users tend to put aside canes and canes' acceptability is reduced. To increase their usability and acceptability, appellative and time-effective technology-based strategies, such as automatic fall detection systems, should be incorporated. First, steps toward a fall-related approach demand the ability to identify that a fall event is eminent relying only on information derived from the device. This study proposes and compares two fall event approaches. A machine learning framework determines the best models and features for three classification problems: fall event, fall phase, and fall category. The long short-term memory (LSTM) is the most suitable classifier for fall event and phase classification problems, requiring the first 32 and 17 features, respectively, ranked by the Relief-F and minimum-redundancy maximum-relevancy methods (accuracy >99% and >95%). K-nearest neighbors with the first 40 features ranked by the Relief-F method presented the best performance for fall category classification (accuracy >74%). A finite-state machine (FSM) combined accelerometer and gyroscope's data to detect the cane's fall with a clear proximity to the LSTM performance (accuracy >99%). Both approaches detect a fall before its occurrence with mean lead times higher than 373 (LSTM) and 266 ms (FSM).