Principal component analysis can decrease neural networks performance for incipient falls detection: A preliminary study with hands and feet accelerations.

Abstract

Fall-related accidents constitute a major problem for elderly people and a burden to the health-care national system. It is therefore important to design devices (e.g., accelerometers) and machine learning algorithms able to recognize incipient falls as quickly and reliably as possible. Blind source separation (BSS) methods are often used as a preprocessing step before classification, however the effects of BSS on classification performance are not well understood. The aim of this work is to preliminarily characterize the effect that two methods, namely Principal and Independent Component Analysis (PCA and ICA) and their combined use have on the performance of a neural network in detecting incipient falls. We used the feet and arms 3D kinematics of subjects while managing unexpected perturbations during walking. Results show that PCA needs to be used carefully as depending on the initial dataset, the PCA might lump variance together thus impairing the performance of an artificial neural networks (ANN) classifier. The use of PCA with 85% residual variance threshold significantly decreased the classifier performance, which was restored with a subsequent ICA (PCA + ICA). The results suggest that BSS techniques, though linear, might have an adverse effect on nonlinear classifiers such as ANN that might be dependent on the initial dataset redundancy.