Linear regression models for chew count estimation from piezoelectric sensor signals

Abstract

Research suggests that there might be a relationship between chewing rate and final energy intake. Wearable sensor systems have been proposed for automatic detection of food intake. This work presents the use of linear regression for estimation of chew counts from piezoelectric sensor signal. For known chewing sequences, four features are computed (number of peaks, valleys, zero crossings and duration of chewing), and linear regression models were trained and tested for estimation of chew counts using cross-validation scheme. Adjusted R2 and mean absolute error (of chew counts) are used for performance evaluation. ANOVA along with Tukey Kramer test was used to compare the performance of different models. Results suggest that best performance was achieved with multiple linear regression model (all features as predictors) with adjusted R2 of 0.95 and mean absolute error of 9.66% ± 6.28%. Results suggest that linear regression models can be used for estimation of chew counts from piezoelectric strain sensor signals.