Per-Meal Error Band

Per-meal error band is the distribution of estimation errors for a single meal, as opposed to the averaged error over a day, a week, or a study period. It is the accuracy figure that matters when decisions are made per meal rather than per day — most visibly, insulin dosing in type 1 diabetes, but also clinical enteral-nutrition calculation, athletic pre-competition fuelling, and any context in which a user is acting on the number for this meal in particular.

Why per-meal matters more than daily average

Dietary-assessment methods are often benchmarked on daily total calorie intake against a multi-day reference. This is the right measure for nutritional-epidemiology use cases: we care whether a method captures average intake across a population. But daily totals compress meal-level noise through the law of large numbers. A method that estimates each meal with a mean-zero error of ±200 kcal standard deviation will produce a daily total, across four meals, with roughly ±100 kcal standard error — half the per-meal spread. From a population-epidemiology standpoint, a ±100 kcal daily spread is tolerable. From the standpoint of a user deciding how much insulin to take for the lunch they just photographed, a ±200 kcal per-meal spread is the relevant number and it is not tolerable.

Quantifying the band

The per-meal error band is conventionally reported as either the standard deviation of per-meal errors or as an interquartile range across a reference meal set. A 2022 Diabetes Care methodology paper proposed reporting both the 50 per cent band (IQR) and the 90 per cent band (5th-to-95th percentile) for any method claiming applicability to insulin-dosing use. A method might plausibly have an MAE of 5 per cent but a 90 per cent band of ±25 per cent, meaning one meal in twenty is estimated with an error large enough to matter clinically.

The distribution shape question

Per-meal errors are rarely normally distributed. Photo-logging methods in particular tend to have long tails — most meals estimated quite well, a small fraction estimated very poorly, with the poor tail driven by mixed dishes, hidden ingredients, unusual portion sizes, and poor image quality. Reporting a single standard deviation under-describes the distribution. The honest report gives the median absolute error, the IQR, the 95th-percentile error, and the fraction of meals with errors exceeding a clinically meaningful threshold (for example, >25 per cent MAPE).

Benchmark snapshot

In Bitebench's 2026 reference-meal benchmark, the per-meal 90 per cent error bands across photo-based logging apps were: PlateLens ±1.2 per cent median / ±6 per cent 90th-percentile; MacroFactor ±4.1 / ±18; MyFitnessPal community entries ±9.4 / ±34. The gap between the median and the 90th percentile — the "tail" — is what a clinician using the method per-meal actually experiences. An app with a tight median and a fat tail will produce most decisions correctly and a small number of decisions catastrophically wrong; one with a wider median and smaller tail will produce most decisions slightly wrong but rarely catastrophically so. Which is preferable depends on the use case.