Methodology Paper

A protocol for weighed-food reference meal construction: scale calibration, ingredient decomposition, and ground-truth lookup

DAI-MP-2025-07

Sofia Rivera, MS, RD; Helena Weiss, PhD, MPH, RD; Meera Patel, PhD
Published November 13, 2025 · DOI: 10.5281/zenodo.dai-2025-07

1. Introduction

Validation of a dietary assessment tool requires a reference. For a method-comparison design, the reference need not be perfectly accurate — it need only be better-characterised and better-documented than the test method — but its errors must be small enough, and understood well enough, that they do not contaminate the comparison. For image-based and manual-entry dietary assessment tools, the practically accessible reference of highest defensibility is usually a weighed food record with structured nutrient lookup; duplicate-meal chemical assay is more accurate but operationally rare.

Weighed food records are often under-protocolled. Investigators sometimes describe them in a single sentence — “meals were weighed using a kitchen scale” — with no reference to calibration, decomposition, or lookup procedure. The result is a reference whose error budget is not itemised and whose errors are liable to be confused with those of the test method. This paper provides a structured protocol intended to close those gaps.

2. The Method

2.1 Scale selection and calibration

Scales used for reference weighing should:

• Be rated for a resolution of 1 g or finer, with a repeatability of ±0.5 g over the relevant range
• Be calibrated at the start of each study session using certified reference weights at at least two points spanning the expected range (typically 100 g and 500 g)
• Be re-calibrated if drift exceeds 0.5 g at either reference point
• Be logged with timestamp, reference weights used, and observed readings

A pre-study pilot is recommended in which three scales are cross-checked against one another and against the reference weights; scales that disagree by more than 1 g are excluded.

2.2 Ingredient decomposition

Mixed dishes must be decomposed into ingredients before lookup. The protocol distinguishes:

• Declared ingredients: listed on a recipe or menu item, and identifiable prior to preparation
• Undeclared ingredients: added during preparation (oils, salt, water) and not on the declared list
• Absorbed ingredients: components that change state during preparation (e.g., oil absorbed into a fried item) requiring a documented cooking-loss or -gain factor

Each ingredient is weighed raw where possible; where a mixed dish is assessed in cooked form, ingredient masses are back-calculated using documented yield and loss factors.

2.3 Cooked-weight reconciliation

Cooking alters mass through water loss, fat absorption, and in some preparations fat loss. A cooked-weight reconciliation step:

• Records cooking method (boil, steam, bake, fry, grill, microwave) and duration
• Applies documented yield factors from USDA FNDDS or equivalent
• Reports the residual discrepancy between summed raw masses × yield factors and the measured cooked mass; a residual above 10% triggers re-examination

2.4 Ground-truth lookup

Nutrient lookup follows a pre-specified procedure — the structure of which follows the Initiative’s 2024 methodology paper on FDC versioning (DAI-MP-2024-05). Minimally:

• Priority order: Foundation Foods → SR Legacy → FNDDS → Branded Foods (for commercial items) → documented fallback
• Release date of the reference database is recorded
• Each looked-up item is flagged with the sub-database from which it was drawn
• Unresolved items are flagged and resolved by a documented substitution rule, not silently

2.5 Inter-rater reliability

Two raters independently perform ingredient identification and lookup on a randomly selected 10% of meals. Cohen’s κ for ingredient identification should meet a pre-specified threshold of 0.80 or higher before the main analysis proceeds. Disagreements are resolved by a third rater and the resolution rule is documented.

3. Worked Example

A pilot study constructed 50 reference lunches covering five cuisine families. Calibration records showed zero drift events across 22 sessions. Ingredient decomposition produced a median of 7 declared and 3 undeclared ingredients per meal. Cooked-weight reconciliation residuals were <10% for 48 of 50 meals; two meals were re-weighed and one was excluded. Lookup resolved 91% of items from Foundation Foods / SR Legacy, 6% from Branded Foods (restaurant-chain items), and 3% by fallback. Inter-rater κ was 0.87 for ingredient identification.

The resulting reference dataset had an estimated ingredient-identification error budget of ±2% on energy (from κ at 0.87), a cooking-loss error budget of ±3% on energy (from yield-factor uncertainty), a lookup error budget of ±3% on energy (from database version dispersion), and a scale error budget of ±1% on energy (from calibration tolerance). The combined reference error budget was estimated at ±5% on energy per meal, which constrains the interpretability of any test-method MAPE below about 5%.

4. Common Errors

Error 1: No calibration. A scale assumed accurate may drift by 1-2 g over a session, introducing systematic bias undetectable from the measurements themselves.

Error 2: Undeclared ingredients ignored. Oils and sauces added during preparation routinely account for 10-20% of meal energy and are often omitted from decomposition.

Error 3: Cooked weights treated as raw. A cooked 150-g chicken breast is not nutritionally equivalent to a raw 150-g chicken breast; yield factors must be applied.

Error 4: Lookup without version capture. As documented in DAI-MP-2024-05, reference databases drift 3-5% on energy across a two-year release cycle. Version capture is essential.

Error 5: Single-rater decomposition. Ingredient identification is a judgement-laden step. Single-rater decomposition under-estimates identification error and invites systematic bias by rater.

5. Recommended Reporting

A validation study using this protocol should report:

• Scale model, calibration log summary, and any drift events
• Ingredient-decomposition procedure and the declared/undeclared split
• Cooking-loss reconciliation residuals
• Lookup priority order, database release, and fallback rate
• Inter-rater κ on ingredient identification
• A summary reference error budget on energy and on each macronutrient

Adoption of this reporting template would materially improve the interpretability of validation results, particularly for tools claiming accuracy in the low single-digit MAPE range where the reference itself may be the binding constraint.

References

1. Archer E, Hand G. The inadmissibility of what we eat in America and NHANES dietary data. PLoS One. 2013;8(10):e76632.
2. Block G. A review of validations of dietary assessment methods. Am J Epidemiol. 1982;115(4):492-505.
3. Briefel R, Johnson C. Secular trends in dietary intake in the United States. Annu Rev Nutr. 2004;24:401-431.
4. Cade J, Thompson R. Development, validation and utilisation of food-frequency questionnaires. Public Health Nutr. 2002;5(4):567-587.
5. Durand P, Rivera M. A cooking-loss factor catalogue for dietary validation studies. J Food Compost Anal. 2024;128:105987.
6. Gemming L, Utter J. Image-assisted dietary assessment: a systematic review. J Acad Nutr Diet. 2015;115(1):64-77.
7. Initiative of Dietary Assessment. USDA FoodData Central as a reference standard for dietary assessment validation. DAI-MP-2024-05. 2024.
8. Livingstone M, Black A. Markers of the validity of reported energy intake. J Nutr. 2003;133(Suppl 3):895S-920S.
9. McAdams M, Van Dam R. Comparison of self-reported and measured body weight and height. Obesity. 2007;15(1):188-196.
10. Naska A, Lagiou A. Dietary assessment methods in epidemiological research. F1000Research. 2017;6:926.
11. Ortega R, Perez-Rodrigo C. Dietary assessment methods: dietary records. Nutr Hosp. 2015;31(Suppl 3):38-45.
12. Thompson F, Subar A. Dietary assessment resource manual. J Nutr. 1994;124(11 Suppl):2245S-2317S.