A Neuro-Fuzzy-Based Sensing Approach for the Classification of Emulated Postural Instability

Abstract

Assistive technology (AT) helps to assess the daily living of frail people and may have a strategic role to detect and prevent falls. In this article, the task of classifying different classes of postural sway behaviors has been addressed by developing a neuro-fuzzy (NF) inference approach that is robust against noise. The proposed approach classifies four different postural behaviors, namely, stable standing (ST), anteroposterior (AP), mediolateral, and unstable (UNST). The strategy exploits data generated by a wearable sensor node, to be positioned on the user chest. A dedicated experimental setup has been realized to emulate the postural dynamics and generate the dataset. Two novel indices to assess the robustness of the system have been proposed. The first index is a measure of residuals between the predicted and the expected postural status, which equally weights estimations with respect to expected classes. The second metric is a reliability index, which allows for assessing the degree of trust of each estimation performed by the NF inference. Results obtained demonstrate the suitability of the proposed methodology, showing a capability of almost 100% to correctly classify patterns among different allowed classes, with reliability indexes of 97.56% and 98.50% for the training and test patterns, respectively. Also, the robustness of the NF classification algorithm against noisy data has been demonstrated.