Investigating Performance of an Embedded Machine Learning Solution for Classifying Postural Behaviors

Postural instability is one of the main critical aspects to be monitored in the case of degenerative diseases, and is also a predictor of potential falls. This paper presents a multi-layer perceptron approach for the classification of four different classes of postural behaviors that is implemented by an embedded sensing architecture. The robustness of the methodology against noisy data and the effects of using different sets of classification features have been investigated. In the case of noisy input data, a reliability index of almost 100% has been obtained, with a negligible drop (less than 5%) being shown for the whole range of noise levels that was investigated. Such an achievement substantiates the better robustness of this approach with respect to threshold-based algorithms, which have been also considered for the sake of comparison.