Predicting Bedside Falls using Current Context

Abstract

Each year about a third of the elderly aged 65 or older experience a fall. Many of these falls could be avoided if fall risk assessment and prevention tools where available in the daily living situation. Such tools would need to use the current context as input to predict an imminent fall. This paper presents an approach predicting imminent falls using data from a roof-mounted infrared array combined with an ultrasonic sensor. The data are processed and features extracted to determine location and posture along with indicators representing movement, direction, and velocity. These features are used by a classification algorithm to create a probability matrix representing the conditional probability of an individual in the current frame being recognized in a specific location and posture. A sequence of these probability matrices are fed into four artificial intelligence constructs trained to predict the probability of a future location/posture. The resulting conditional probability is used as a fall risk indicator to predict falls. Finally, the results from the experiment are presented. The study concludes that Elman Recurrent Neural Network with adapted Teacher Forcing has very promising properties and an explanation of the findings is offered.