Poster Topical Area: Nutritional Epidemiology
Location: Hall D
Poster Board Number: 754
Objectives: Monitoring caloric intake is becoming necessary as the number of pathologies related to obesity and overweight increase . New wearable devices may help people to automatically record the nutrition intake. Wrist banded triaxial accelerometer has been proposed as potentially useful device for the purposes of monitoring caloric intake  but never for food classification.Instead, food classification during an eating occasion is important for monitoring diet adherence in all unstructured context (like restaurant or out-of-home meals). This study aims to develop more robust and effective methods for the food intake classification from data obtained when eating.
Methods: A supervised pilot study was conducted on 20 healthy young people (10 women and 10 men aged 20-31 years). Each subject was equipped with two wearable devices (Garmin Fenix 5) for each wrist, and seven food items have to be eaten under semi-naturalistic conditions. All meals were registered and acceleration data were then extracted and analysed. We added two other time series, the magnitude of the acceleration and decomposition into principal components (PCA) . The features proposed in this paper are quite popular among the experts in this field [4, 5]. Mean, standard deviation, energy, correlation and entropy were extracted from acceleration data. Food recognition on these features was performed using several classifiers. We evaluate the performance of the classifiers by traditional means of splitting training and testing data.
Results: Overall, recognition accuracy is highest for Weighted k-Nearest Neighbors with an accuracy rate of 96% (CI 95%,97%), compared to 95% (CI 94%,96%) achieved by Support Vector Machine and Random Forest, and 90% (CI 88%,92%) achieved by Bagging.
Conclusions: Our pilot work shows that monitoring food items via the usage of simple wrist-banded wearable devices is feasible and accurate. Machine learning tools are necessary to deal with the complexity of signals gathered by the devices, and research is ongoing on how to (i) improve further accuracy, (ii) considering the interaction among both wrists and other devices (like mobiles) and (iii) work on large scale implementation and testing.
 World Health Organization et al. Noncommunicable diseases progress monitor 2015. 2015.
 Dario Gregori, Clara Minto, Corrado Lanera, and Giulia Lorenzoni. Feasibility and reliability
of wearable devices in measuring caloric intake: Results from a pilot study. The FASEB Journal,
31(1 Supplement):302–6, 2017.
 Andrea Mannini and Angelo Maria Sabatini. Machine learning methods for classifying human
physical activity from on-body accelerometers. Sensors, 10(2):1154–1175, 2010.
 Ling Bao and Stephen Intille. Activity recognition from user-annotated acceleration data. Pervasive
computing, pages 1–17, 2004.
 Nishkam Ravi, Nikhil Dandekar, Preetham Mysore, and Michael L Littman. Activity recognition
from accelerometer data. In Aaai, volume 5, pages 1541–1546, 2005.
Unit of Biostatistics, Epidemiology and Public Health, Department of Cardiac, Thoracic andVascular Sciences, University of Padova, Italy
Maserà di Padova, Veneto, Italy