Aristotle University Of Thessaloniki (AUTH)
Modeling and automatically measuring human eating behavior
Abstract
dc:descriptionThis work proposes models and algorithms that leverage signals collected from wearable sensors to objectively quantify human eating behavior by extracting eating-behavior indicators. We focus on two different sensors that capture different, yet complimentary, aspects of eating behavior and propose new algorithms for each sensor to derive these indicators, focusing mainly on the generalization under different situations such as noisy background and usage during real-life activities. The first device is a prototype multi-modal sensor that captures audio, photoplethysmography, and acceleration signals. We explore some properties of audio signals of chews, such as their fractal dimension, and propose algorithms based on feature extraction and classification with support vector machines, as well as end-to-end approaches with convolutional neural netwroks. We also propose (for the first time in literature) the use of photoplethysmography as a means to detect chewing, both independently and combined with audio. We propose and evaluate different algorithms. The best algorithm uses a support vector machine classifier with spectral features and is also combined in a late-fusion scheme with a respective audio-based algorithm. This late-fusion approach achieves higher effectiveness compared to each sensor individually and can be improved even further by taking into account the energy signal from a belt-mounted three-dimensional accelerometer (F1-score of 0.76 for a leave-one-subject-out experiment on a challenging, real-life dataset). The audio sensor is also used to recognize food attributes such as crispiness. We form a multi-label problem where labels correspond to food attributes and propose recognition algorithms both for individual chews and for chewing bouts. Experiments show that it is possible to generalize these properties to new, unknown (to the training stage) food types with high effectiveness, in some cases (0.92 weighted accuracy per bout for a leave-one-food-type-out experiment). The second sensor continuously measures the weight of the food inside a plate during a meal. Based on the captured weight signal we can derive useful in-meal indicators of eating behavior. We propose two types of algorithms. The first type focuses on identifying moments when additional food is placed on the plate, since these moments change the generally decreasing trend of the captured weight (as food is removed from the plate to be consumed). The effect of adding food is then removed from the captured signal and bites are identified as abrupt weight decrements. The second type of algorithm models the events of a meal with the help of a context-free grammar. Maximum-likelihood is then used to select the most likely parse tree of the meal. The parse tree is an interpretation of the meal structure and events, and is used to derive the in-meal indicators. The context-free grammar and maximum-likelihood--based algorithm is the one who achieves the lowest average (per meal) error rates for almost all indicators (e.g. 24 grams for total food intake, and 1 minute for meal duration).
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2019
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Papapanagiotou, Vasileios
- Παπαπαναγιώτου, Βασίλειος
Subjects
dc:subject × 12- Συμπεριφορά σχετική με βρώση
- Φορετοί αισθητήρες
- Μηχανική μάθηση
- Ψηφιακή επεξεργασία σήματος
- Eating behavior
- wearable sensors
- Machine learning
- Digital signal processing
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
dc:rights- Statement dc:rights
-
- BY_NC_ND
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/46594