Novelty detection for predicting falls risk using smartphone gait data

In this paper, we consider the problem of falls risk prediction in elderly adults using smartphone-based inertial gait measurements. We begin by collecting a parallel data set from a pressure sensitive walkway and smartphones. The walk-way data is used to calculate the falls risk ground truth using well-established biomechanical norms. The smartphone data and falls risk labels are then used to train and evaluate both the one-class support vector machine (OC-SVM) and the support vector data description (SVDD) novelty detectors. In our evaluation, we find the SVDD has an average F
₁
score, used as a measure of classifier performance by equally weighting precision and recall, of 76% for females and 79% for males compared to 79% for a universal model. These results demonstrate the potential for predicting falls risk from smartphone data using novelty detection.