Probabilistic parsing of dietary activity events

Abstract

Dietary behaviour is an important lifestyle aspect and directly related to long-term health.We present an approach to detect eating and drinking intake cycles from body-worn sensors. Information derived from the sensors are considered as abstract activity events and a sequence modelling is applied utilising probabilistic context-free grammars. Different grammar models are discussed and applied to dietary intake evaluation data. The detection performance for different foods and food categories is reported. We show that the approach is a feasible strategy to segment dietary intake cycles and identify the food category.