Prediction of Elderly Falls Using the Degree of Cyclostationarity of Walk Pressure Signals

There is an increasing interest in developing older adult fall-risk prediction models that can be used as a preventive approach to predict future risk of falling in the elderly community. This study's primary objective is to implement and compare supervised machine-learning methods to classify elderly subjects as fallers or non-fallers. Features used for building the models were extracted from the pressure signals of the innersoles of 520 elderly people who reported whether they had experienced previous falls or not. Two different types of feature sets were used as inputs to the classification models and were compared. The first feature set includes ten time-domain features, while the second set includes a single cyclostationary property, which is the degree of cyclostationarity of the average walk pressure signal. Our study showed that the use of the degree of cyclostationarity as a single feature improved the model prediction accuracy by 6.58%, compared to the use of the time-domain features. The results obtained prove that cyclostationary features are essential features for the development of classification models for identifying fallers in the elderly community.