Machine Learning Approach to Classify Postural Sway Instabilities

Wearable sensing devices have been extensively proposed for monitoring frailty subjects' mobility and related risk of falls. Considering the key-value of instability to assess degenerative diseases such as Parkinson's disease and its impact on the life quality for this class of end-users, reliable solutions that enable a continuous and real time estimation of postural sway might play a fundamental role. In this paper a machine learning approach to classify among 4 different classes of postural behaviors (Standing, Antero-Posterior sway, Medio-Lateral sway, Unstable) is investigated. The classification algorithm is compliant with its implementation in the adopted embedded architecture, which is equipped with sensors and an Artificial Intelligence core. The proposed approach demonstrates suitable performances in terms of accuracy in correctly classifying unknown patterns as belonging to the right postural sway class. An accuracy index higher than 98% and a very promising reliability index better than 98% have been obtained. The robustness of the algorithm with respect to the dataset organization has been also assessed, and a comparative analysis against threshold-based approaches is also presented.