Novel recurrence features for prefall and fall detection in backward and forward fall types

Injuries caused by different types of falls are one of the vital health threats to the elder community living independent or otherwise. Characterization and detection of a fall event can trigger an alert and minimize the damage. This work presents recurrence quantification parameters as novel features for characterization of a fall event in case of backward and forward types of falls using data acquired through wearable sensors. Computing cross recurrence plots and recurrence parameters; recurrence rate (RR), determinism (DET) and line entropy (ENT) for pre-fall, fall and post-fall phases, the level of signal stability and non-stability is quantified. The recurrence parameters show a stable behaviour in case of pre-fall (RR=0.74, DET=0.85, ENT=4.36) and chaotic behaviour in case of fall (RR=0.39 DET=0.80, ENT=3.13). To assess the discriminating capability of novel recurrence features, a support vector machine (SVM) is used to perform binary classification for prefall and fall classes. The SVM results in overall accuracy of 76% with a positive prediction of 82% for fall and 70% for pre-fall events. The results indicate that recurrence metrics are successfully able to characterize a sudden fall event and could be used in designing fall detection algorithms using wearable sensors.