AI-powered calorie monitoring apps can estimate a meal’s dietary content material from a single {photograph}. The expertise presents a quick and handy various to manually getting into each meals and portion, however new analysis suggests the outcomes could also be significantly decrease than what is definitely on the plate.
In a check of 4 photo-based apps, researchers discovered that calorie and fats estimates have been about one-third too low on common.
How AI Estimates Energy From Meals Pictures
Photograph-based calorie monitoring depends on AI picture recognition to establish the meals proven in an image and estimate the scale of every portion. The app then compares these estimates with diet databases to calculate energy and different vitamins.
“Photograph-based calorie monitoring apps are very fashionable, particularly for individuals attempting to handle their well being or shed weight,” stated Aaron Hengist, a postdoctoral visiting fellow with the Intramural Program of the Nationwide Institute of Diabetes and Digestive and Kidney Ailments (NIDDK), a part of the Nationwide Institutes of Well being. “Nonetheless, the accuracy of many of those apps has not been completely evaluated. Our examine helps tackle this query by taking a look at whether or not these apps can reliably estimate energy.”
Olivia Charles, a postbaccalaureate intramural analysis coaching fellow at NIDDK, introduced the findings at NUTRITION 2026, the flagship annual assembly of the American Society for Diet, held July 25-28 in Nationwide Harbor, Maryland, simply outdoors Washington, D.C.
Testing Apps With Exactly Measured Meals
The challenge is a part of a broader diet examine on the NIH Scientific Heart that’s investigating how the physique processes vitamins on both a low-carbohydrate (ketogenic) weight loss plan or a regular weight loss plan.
Meals used within the scientific trial are ready in a managed metabolic kitchen, the place researchers measure components to the closest 0.1 gram. This gave the workforce a extremely correct reference for evaluating the apps.
Researchers collected standardized images of 102 meals ready for the weight loss plan examine. They then submitted the photographs to MyFitnessPal, LoseIt!, CalAI and Appediet to see how intently every app’s estimates matched the recognized dietary content material.
“By utilizing meals ready in a tightly managed metabolic kitchen, we have been in a position to evaluate the apps’ estimates in opposition to a exact reference,” stated Hengist. “This type of direct, high-quality comparability hasn’t been obtainable earlier than.”
Apps Missed Lots of of Energy
Throughout all 4 apps, estimated calorie totals have been about 250 to 345 energy too low per meal on common. The apps additionally underestimated fats by roughly 30 grams.
MyFitnessPal and LoseIt! have been extra correct when analyzing higher-calorie meals than they have been with lower-calorie meals. All 4 apps additionally produced extra constant estimates for carbohydrates than for different macronutrients.
“Individuals utilizing a photo-based monitoring app with out adjusting the parts or getting into the quantities of meals ought to take the outcomes with a grain of salt,” stated Hengist. “These apps are likely to underestimate energy, particularly from fat, so what they really ate is probably going increased than what the app reveals.”
Keto Meals Could Be More durable for AI to Measure
Following the primary evaluation, the researchers examined greater than 200 further meals to analyze which elements would possibly affect app accuracy.
Preliminary findings point out that the apps might have better problem evaluating meals from a low-carb ketogenic weight loss plan. These meals usually comprise extra fats, which the apps tended to underestimate persistently.
The researchers recommend that combining photo-based instruments with conventional strategies of evaluating meals consumption and weight loss plan high quality may make calorie monitoring extra correct in on a regular basis use.
Charles introduced this analysis on Saturday, July 25, throughout the President’s Oral Session within the Gaylord Nationwide Resort & Conference Heart (summary).
Abstracts introduced at NUTRITION 2026 have been reviewed and chosen by a committee of consultants. Nonetheless, they’ve usually not accomplished the total peer evaluate course of required for publication in a scientific journal. The outcomes ought to subsequently be thought-about preliminary till they seem in a peer-reviewed publication.
