A measurement strategy to assess performances of Fall Detector paradigms

Abstract

The high rate of falls in frail people, represent a serious issue which have catastrophic consequences. To reduce the society impact of such phenomenon, researchers and physicians, have tried to develop solutions ranging from gait training to more technological approaches such as the assistive device. Assistive device are typically focused to fall and ADLs detection and classification. Although two are the main computing approaches for fall detection, i.e threshold based algorithms and machine learning methods, up today a single and well-accepted solution has not been yet identified.In this paper a measurement based methodology to assess performances of measurement systems for Fall Detection is presented. A case of study is discussed to demonstrate the suitability of the strategy proposed in case of a fall detector exploiting inertial sensors and a NeuroFuzzy paradigm.The aim of the assessment approach is twofold: i) to test the system robustness against learning and test patterns combinations; ii) to assess the system performances in terms of reliability in fall classification. The strategy makes a large use of metrics such as sensitivity and specificity which are essential to quantify the goodness of an algorithm. Although this approach has been developed for a machine learning method, it is also suitable for other classification techniques. Different synthetic indexes have been provided, quantifying the overall performances of the classifier are respect to test patterns combination and considered events. As regard the Sensitivity an average value of 93,41 has been measured while the Specificity has a value of 98,06.