A Human-Mimetic AI System for Automatic, Passive and Objective Dietary Assessment - A Human-Mimetic AI System for Automatic, Passive and Objective Dietary Assessment
Unhealthy diet is strongly linked to risks of chronic diseases, such as cardiovascular diseases,
diabetes and certain types of cancer. The Global Burden of Disease Study has found that, among
the top 17 risk factors, poor diet is overwhelmingly the No. 1 risk factor for human diseases.
Despite the strong connection between diet and health, unhealthy foods with large portion sizes
are widely consumed. Currently, 68.5% of U.S. adults are overweight, among the highest in
developed countries. The recent decline in U.S. life expectancy sent another alarming signal
about the general health of the American people. Understanding how the diet-related risk factors
affect people’s health and finding effective ways to empower them in improving lifestyle habits are
among the most important tasks in public health. Unfortunately, dietary assessment in real-world
settings has been exceedingly complex and inaccurate to implement. Technology is needed that
allows researchers to assess dietary intake easily and accurately in real world settings so that
effective intervention to manage obesity and related chronic diseases can be developed. We
propose a biomedical engineering project to address the dietary assessment problem,
taking advantage of advanced mathematical modeling, wearable electronics and artificial
intelligence.
Our research team has been improving the ability to assess diet for over a decade. We have
designed the eButton, a small wearable device pinned on clothes in front of the chest, capable of
collecting image-based dietary data objectively and passively (i.e., without depending on subject’s
self-report or volitional operation of the device). We have also developed algorithms to compute
food volumes and nutrients from images. Since the eButton was developed, it has been used by
many researchers in the U.S. and other countries for objective and passive diet-intake studies in
both adults and children.
Despite the past successes, there have been two lingering critical problems associated with
the objective and passive dietary assessment using wearable devices: 1) substantial manual
efforts are required for researchers to visually examine image data to identify foods and estimate
their volumes (portion sizes), and 2) there are privacy concerns about researchers’ viewing of
participants’ real-life images. Although solving these problems could enable the eButton and other
wearable devices for large-scale diet-intake studies, we were not able to find effective solutions
until recently when Artificial intelligence (AI) emerged. Advanced AI systems, especially those
based on deep learning, can be trained by large amounts of labeled data to produce results
comparable or even superior to those produced by human in numerous fields of applications. AI
technology is also a powerful tool for dietary assessment, potentially providing an ideal solution
to the two previously mentioned problems. We thus propose to develop a human-mimetic AI
system to recognize foods from images, estimate portion sizes, and find energy and nutrient
values from a database in a fully automatic process. Using the AI approach, there will be no need
for researchers to view participants’ real-life images, and the AI system well-respects individuals’
privacy because it is trained to recognizes human foods only, nothing else.
Currently, the performances of existing AI systems are limited by the extensive variety and
high variability of human foods, insufficient training data, and difficulty in finding appropriate
nutritional information from food databases. In this application, we propose a new strategy to
personalize the AI system for each research participant using an advanced mathematical model
of personal food choices. With this personalization step, the dimensionality of our envisioned AI
system can be reduced drastically, and our goal of automatic, objective and passive dietary
assessment can be reached realistically. We also propose to improve the electronic hardware
and develop a biomimetic camera to enlarge the field of view for the eButton. Finally, we will
conduct a thorough evaluation of the personalized AI system in real-world settings using human
subjects.
