Study
How close is a Healthier calorie estimate?
On 250 weighed meals, Healthier’s calorie estimates were within 10% three times out of four, and within 20% nineteen times out of twenty. The average miss was 48 kcal on a 612 kcal meal. Mean absolute error was 8.4%, with a 95% confidence interval of 7.6–9.3%.
Georg Widschwendter and Georg Hirschi, Healthier. A Healthier study of weighed meals.
Download the validation report (PDF)
What was tested
Each reference meal was built from weighed ingredients, a cooked-yield adjustment, a weighed served portion, and verified food-composition data. The person running Healthier did not know those numbers. The app, model, and nutrition data were locked, and only the first estimate counted. The 250 meals cover simple plates, home cooking, restaurant-style meals, packaged food, snacks, and drinks.
Results
Primary endpoint, 250 meals. Mean reference meal 612 kcal (range 118–1,420).
| Metric | Result |
|---|---|
| Meals in the primary set | 250 |
| Mean absolute percentage error | 8.4% |
| 95% confidence interval for MAPE | 7.6–9.3% |
| Median absolute percentage error | 7.1% |
| Mean absolute error | 48 kcal |
| Root mean squared error | 71 kcal |
| Mean signed percentage error | −1.8% |
| Meals within 10% | 74% |
| Meals within 20% | 94% |
Confidence interval from 10,000-sample bootstrap, as specified before the test.
How other errors are usually described
Other published figures, for context. They used different meals and different methods.
- Accredited sport nutritionists estimating meals from photos were about 14–18% off on energy for complex meals.
- Everyday MyFitnessPal logs missed by hundreds of calories a day, largely from forgotten foods and portion entry. Under controlled entry the gap is smaller, and portion size is still the hard part.
- Typing a meal into a database depends on picking the right food and the right portion.
- A 2025 study of ChatGPT-4 on 114 meal photos found about 27% average error on nutrients and portion weight.
- A 2023 review of photo-based dietary assessment reported calorie errors from under 1% to about 38%, depending on the method.
- A US nutrition label can still be in compliance when a lab test is up to 20% above the printed calories.
- EU guidance allows larger tolerances for some nutrients, including vitamins and minerals.
A note on real meals
The study used clear, weighed meals. A rushed photo or a hidden sauce can move the number, and you can correct any item before you save. The figure is the first estimate, which is the one that has to be good.
Questions
How was this tested?
250 meals were weighed, cooked, plated, and weighed again. Healthier estimated each one without seeing the scale weights. Only the first estimate counted.
What does 8.4% mean for a meal?
On a 612 kcal meal, the average miss was 48 kcal. Three in four meals were within 10%. Nineteen in twenty were within 20%.
How can a photo be that close?
Healthier names the foods it can see and estimates the portion. The study measured that first answer against a weighed plate, before anyone edited it.
When should I correct an estimate?
When oil, a sauce, or a drink is easy to miss. Say it, or scan the barcode on a package. The saved meal updates immediately.
Does this work for restaurant meals?
Yes. Restaurant-style meals were part of the 250. Hidden ingredients are the usual reason to add a word or two.
How does this compare with typing food into a database?
Typing depends on finding the right entry and guessing the portion. Published tests of that workflow show much larger everyday gaps. Healthier’s number is from weighed plates, not from a typing test.
What about protein and micronutrients?
The study’s endpoint was calories. Protein, carbs, fat, and micronutrients are estimated from the same items in the meal.
How should I describe the result?
“On 250 weighed meals, Healthier’s calorie estimates were within 10% three quarters of the time (mean error 8.4%).” Link this page.