Empirical mode decomposition of throat microphone recordings for intake classification

Abstract

Wearable sensor systems can deliver promising solutions to automatic monitoring of ingestive behavior. This study presents an on-body sensor system and related signal processing techniques to classify different types of food intake sounds. A piezoelectric throat microphone is used to capture food consumption sounds from the neck. The recorded signals are firstly segmented and decomposed using the empirical mode decomposition (EMD) analysis. EMD has been a widely implemented tool to analyze non-stationary and non-linear signals by decomposing data into a series of sub-band oscillations known as intrinsic mode functions (IMFs). For each decomposed IMF signal, time and frequency domain features are then computed to provide a multi-resolution representation of the signal. The minimum redundancy maximum relevance (mRMR) principle is utilized to investigate the most representative features for the food intake classification task, which is carried out using the support vector machines. Experimental evaluations over selected groups of features and EMD achieve significant performance improvements compared to the baseline classification system without EMD.