This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Research Protocols, is properly cited. The complete bibliographic information, a link to the original publication on http://www.researchprotocols.org, as well as this copyright and license information must be included.
The monitoring of caloric intake is an important challenge for the maintenance of individual and public health. The instruments used so far for dietary monitoring (eg, food frequency questionnaires, food diaries, and telephone interviews) are inexpensive and easy to implement but show important inaccuracies. Alternative methods based on wearable devices and wrist accelerometers have been proposed, yet they have limited accuracy in predicting caloric intake because analytics are usually not well suited to manage the massive sets of data generated from these types of devices.
This study aims to develop an algorithm using recent advances in machine learning methodology, which provides a precise and stable estimate of caloric intake.
The study will capture four individual eating activities outside the home over 2 months. Twenty healthy Italian adults will be recruited from the University of Padova in Padova, Italy, with email, flyers, and website announcements. The eligibility requirements include age 18 to 66 years and no eating disorder history. Each participant will be randomized to one of two menus to be eaten on weekdays in a predefined cafeteria in Padova (northeastern Italy). Flows of raw data will be accessed and downloaded from the wearable devices given to study participants and associated with anthropometric and demographic characteristics of the user (with their written permission). These massive data flows will provide a detailed picture of real-life conditions and will be analyzed through an up-to-date machine learning approach with the aim to accurately predict the caloric contribution of individual eating activities. Gold standard evaluation of the energy content of eaten foods will be obtained using calorimetric assessments made at the Laboratory of Dietetics and Nutraceutical Research of the University of Padova.
The study will last 14 months from July 2017 with a final report by November 2018. Data collection will occur from October to December 2017. From this study, we expect to obtain a series of relevant data that, opportunely filtered, could allow the construction of a prototype algorithm able to estimate caloric intake through the recognition of food type and the number of bites. The algorithm should work in real time, be embedded in a wearable device, and able to match bite-related movements and the corresponding caloric intake with high accuracy.
Building an automatic calculation method for caloric intake, independent on the black-box processing of the wearable devices marketed so far, has great potential both for clinical nutrition (eg, for assessing cardiovascular compliance or for the prevention of coronary heart disease through proper dietary control) and public health nutrition as a low-cost monitoring tool for eating habits of different segments of the population.
DERR1-10.2196/12116
Obesity affects 650 million people worldwide and is strongly related to cardiovascular diseases, which are the main cause of death in Western countries [
With advances in technologies, studies have started to employ electronic wearable devices for monitoring caloric intake [
Wearable devices equipped with motion sensors usually detect feeding-related gestures, translate gestures into a number of bites, and then bites into energy intake through the use of algorithms whose rationale is based on the existence of a relationship between the number of bites and the number of calories eaten [
The main reason they are poor in associating bites to calories is because of their poor analytics, which are usually not well suited to manage the massive sets of data generated from these devices and with a limited accuracy in predicting caloric intake [
This section describes the protocol of the study “Measuring Caloric Intake at the Population Level” (acronym NOTION), approved by the Bioethics Committee of the University of Torino, Torino, Italy, on July 12, 2017 (#256091).
The study will last 14 months from July 2017 with a final report by November 2018. Data collection will occur from October to December 2017.
The study will capture four individual eating activities outside the home over a 2-month period. Twenty healthy Italian people will be recruited from the University of Padova with email, flyers, and website announcements. The eligibility requirements include age 18 to 66 years, no eating disorder history, and absence of allergies and food intolerances. Potential participants will complete a Web-based initial form to collect a preliminary assessment of the fulfillment of the inclusion criteria. Those who appear eligible will meet with the research team members who will describe the study, answer questions, and obtain written informed consent to determine final eligibility. Participants will receive an eating activity schedule, a device, and written instructions.
Each participant will be required to eat a minimum of four meals on weekdays in a predefined cafeteria in Padova (northeastern Italy).
To collect information about dietary habits and physical activity, a standard questionnaire will be administered to all enrolled participants [
A selection of different types of prepackaged foods will be made by dieticians to simulate a complete Mediterranean daily diet of 2000 kcal or less. The choice of prepackaged food is driven by the fact that bias regarding the energy content will be reduced because the caloric information reported on the labels will be compared to the values measured by the calorimeter. Two menus (menus A and B,
Two digital cameras will be used to record each participant’s mouth, torso, and tray during meal consumption. Four Garmin Fenix 5 watches containing an accelerometer and gyroscope will be used to record the movement of the wrists of each participant at 5 Hz. The FIT files generated by the watches will be converted using the java/FitToCSV.bat utility in the FIT SDK. They will contain all the relevant information for the analysis: triaxial accelerometer data x, y, z, pitch and roll angle, power, and time.
Gold standard evaluation of the energy content of selected foods will be obtained by calorimetric assessments made at the Laboratory of Dietetics and Nutraceutical Research of the Department of Cardiac, Thoracic and Vascular Sciences of the University of Padova, Padova, Italy.
For this purpose, all collected bites will be measured by mass, while two bites of each food item per participant will be randomly selected to measure volume and energy content. The mass of the bites will be measured using the Gibertini Crystal 100 scale. The volume measurement will be obtained measuring the liquid displacement due to the immersion of the sample in a graded cylinder (250 mL class A) filled with water. Caloric content of bites will be measured using a bomb calorimeter after the complete homogenization of a single bite [
Composition of the two menus (menus A and B) randomly assigned to the volunteers recruited in the NOTION study.
Menu and meal | kcal/portion | ||
Rusks and jam | 114 kcal/41 g | ||
Yogurt | 186 kcal/170 g | ||
Risotto with asparagus | 471 kcal/300 g | ||
Mozzarella | 242 kcal/100 g | ||
Biscuits | 278 kcal/55 g | ||
Vegetable soup | 135 kcal/310 g | ||
Artichoke chicken | 199 kcal/120 g | ||
Brioche | 99 kcal/28 g | ||
Yogurt | 186 kcal/170 g | ||
Tagliolini with mushrooms | 546 kcal/300 g | ||
Chicken meatballs with tomato sauce | 135 kcal/120 g | ||
Sandwich | 195 kcal/70 g | ||
Eggplant parmigiana | 309 kcal/300 g | ||
Italian fresh cheese | 269 kcal/100 g |
NOTION study design and flowchart.
The calorimeter calculates the gross energy of the samples; therefore, to consider the metabolizable energy it will be necessary to correct the value by subtracting 1.25 kcal per gram of protein from gross energy.
Each volunteer will be assigned a menu through randomization, providing two balanced groups of 10 participants for each menu. A sampling strategy will be applied to select 2 of 10 bites collected during the simulated meal for chemical and physical analysis.
With the aim to randomize each participant to a menu, the “blockrand” package of R will be used. This function creates random assignments for clinical trials, or any experiment, in which subjects are enrolled sequentially. Blocked randomization with a 1:1 ratio will be performed to ensure balance between the groups throughout the study.
Bites per food item and participant will be numbered progressively from 1 to 10 as they are collected. After random reordering, performed with the “sample” function of R, the first two will be sampled. Since sampling is scheduled in advance, and a participant could complete the meal in less than 10 bites, if the bite indicated in the reordered list be unavailable, then the next in the list will be considered (eg, assume that the list is 1,10, 3, 2, 7, 4, 8, 5, 9, 6 and the participant has finished the meal in five bites, then the first and third bites will be analyzed).
A human rater will annotate an Excel file by watching the video and pausing it at times when a bite is seen to be taken, using frame-by-frame rewinding to identify the time when food or a beverage is placed into the mouth. Two human raters independently will label each food. Raters will be trained to standardize the process of labeling.
To achieve synchronization without wires of cameras and wrist motion trackers, device clocks will be synchronized with the participants’ watch times at the beginning of each data collection session. Participants will be asked to clap their hands at the beginning of each eating session. Watch times will be manually recorded for the periods of clapping. The peaks of the distinct sinusoidal patterns at the beginning of each acceleration signal will be visualized on the video and aligned between the wearable devices.
Working with sensor signals coming from wearable devices that monitor human movements involves the use of signal processing techniques to remove noise from the data [
Row acceleration signals consist of movement, gravitational, and noise components. Separation of these becomes difficult during rotational movements [
The use of accelerometer data has recently emerged in several biomedical apps (eg, [
The purpose of this study is obtaining an estimate of caloric intake from the movements of the wrist. To achieve this objective, we will exploit accelerometer data because it is known to be a suitable adoption for this kind of problem [
Building an automatic calculation method for caloric intake independent of the black-box processing of the wearable devices marketed so far, has great potential both for clinical nutrition (eg, for assessing cardiovascular compliance or for the prevention of coronary heart disease through proper dietary control) and public health nutrition as a low-cost monitoring tool for eating habits of specific segments of the population.
From this study, we expect to obtain a series of relevant data that, opportunely filtered, could allow the construction of a prototype algorithm able to estimate caloric intake through the recognition of food type and the number of bites. The algorithm should work in real time and be embedded in a wearable device able to match bite-related movements and the corresponding caloric intake with high accuracy. The challenge will be refining the algorithm to recognize an extended number of foods on a large sample of multiethnic participants.
The estimation of caloric intake is a topic of great interest in the field of nutrition [
To handle the challenges of big data from wearable devices, new statistical thinking and computational methods are needed. The use of machine learning techniques, able to grab all the available information to solve complex learning tasks such as classification, clustering, and numerical prediction, represent a possible advancement.
This study will address these existing challenges and make the following innovative contributions:
Build a detailed and objective picture of eating activity under seminaturalistic conditions, as a reliable and robust basis for the subsequent analytical steps.
Use cutting-edge advances in machine learning for big raw data generated by inertial sensors.
Implement a prototype algorithm for caloric intake, independent on the wearable device used at the development stage, which will hopefully improve dietary monitoring.
We acknowledge that the whole system is still premature for real-time implementation and that future work will require a broader range of food items, a larger sample size, and a free-living condition. Furthermore, we will need to check the hypothesis that the type of food rather than other participant and food characteristics (eg, cutlery used to eat, anthropometry) is the best proxy of caloric intake. Moreover, we will have to investigate whether the signals recorded by the wearable devices actually hold promise in measuring caloric intake. Nevertheless, this ambitious project has great potential to empower dietary monitoring.
The NOTION study was supported by funding of the Italian Ministry of Health, Directorate-General for hygiene, food safety and nutrition.
The following are members of the NOTION Group: Alessandra Corazzina, Gianluca N Piras, Loro Christian, Lanera Corrado, Valentini Romina, Caregaro Lorenza, and Baldas Solidea.
None declared.