Ground Truth
Ground Truth. The best available value for what was actually measured, used as the yardstick an instrument's readings are checked against.
The yardstick, not the thing itself
Ground truth is the value you treat as correct when you test an instrument. The term comes from remote sensing, where satellite readings were checked against measurements taken on the ground, but the idea is general: to know how wrong something is, you need something you trust more.
The important word is trust more, not perfect. Ground truth is itself a measurement, with its own uncertainty. A good validation picks a reference that is much more accurate than the instrument under test and says how accurate it is.
Options for calorie intake
Different questions call for different references:
| Reference | What it gives you | Typical use |
|---|---|---|
| Doubly labelled water | Total energy expenditure over one to two weeks, which equals intake when weight is stable | Checking whether self-reported intake over a period is plausible |
| Chemical analysis of a duplicate meal | Measured energy of a portion identical to the one eaten | The most direct per-meal reference, but costly |
| Weighed components plus composition data | Calculated energy of a specific meal | The practical standard for testing apps meal by meal |
| A trained dietitian’s estimate | An expert judgement | Useful when nothing else is possible, but not a true reference |
For calorie-tracking apps, the third row is the usual choice, because the question is per-meal and it can be repeated across many meals at a reasonable cost.
Where the composition values come from
In the United States the standard source is USDA FoodData Central, which contains several datasets with different origins:
- Foundation Foods: values from recent laboratory analyses, with detailed documentation.
- SR Legacy: the former Standard Reference database, analytical values from earlier decades.
- Branded Foods: label information supplied by manufacturers and data providers.
- FNDDS: values for foods as reported in national dietary surveys.
A careful protocol prefers Foundation Foods, falls back to SR Legacy, and uses Branded Foods only for packaged products, checked against the label in hand. Outside the US, national tables play the same role for local foods; for non-US dishes we use peer-reviewed regional databases (UK COFID, the Indian National Institute of Nutrition database, FAO compositions) where appropriate.
Worked example, and its uncertainty
A test meal is 150 g of cooked white rice, 120 g of roasted chicken breast and 10 g of olive oil. With USDA values of about 130, 165 and 884 kcal per 100 g, the ground truth is 195 + 198 + 88 = 481 kcal.
That number is a best estimate, not an exact fact. The chicken breast on your plate is not the average chicken breast in the database; a particular banana may differ from the FoodData Central banana entry by 5-10%. The scale adds a little more uncertainty, especially for small amounts of oil. A well-run study accepts this because the reference error is small compared with the errors it is trying to detect in apps, and it reports how ground truth was built so others can judge.
Why it matters when you read accuracy claims
An accuracy figure is only as meaningful as its ground truth. “Accurate to within 5%” against weighed meals and composition data is a real measurement. “Accurate to within 5%” against the app’s own database, or against what users said they ate, measures almost nothing. Before trusting any figure, find out what it was measured against. The weighed portion reference entry explains how the practical standard is built, and the MAPE entry how error against it is summarised.