Food Types

Our database distinguishes three food types:

• Generic single foods
• Generic composite foods
• Branded foods

Terminology

“Generic” foods represent a product category by the mean of a sample. As a consequence, they do not capture product-specific differences (e.g., fortification, added sugar, salt reduction).

Generic Single Foods

Generic single foods consist of exactly one ingredient (e.g., “apple”). They are based on a manually curated food list that initially includes only raw-form products with nutrient values from the Bundeslebensmittelschlüssel 4.0 (BLS 4.0). Each entry is assigned a master category that standardizes available processing states.

Development outlook

In the future, the master category will be replaced by hierarchical FoodEx2 coding to improve the synonym database and support upcoming features.

Following the mHEI methodology, derived foods (e.g., “apple juice”) are linked to their parent food (e.g., “apple”) to ensure correct calculation of mHEI equivalents. For processing states, nutrient changes are calculated using yield factors and retention factors from the scientific literature.

Gaps in the BLS food selection

BLS 4.0 lacks some foods (e.g., “hemp seeds”). We continuously close these gaps with foods whose nutrients are derived from the most similar BLS item.

Generic Composite Foods

Generic composite foods are recipes (e.g., “apple spritzer”) made entirely from generic single foods with specified quantities and processing states. This allows their nutrient profile to be calculated directly and precisely. The selection includes household recipes (e.g., “waffle batter”), loose goods (e.g., “bread roll with meatloaf”), and generic alternatives for branded foods (e.g., “cola beverage with caffeine”).

Tip: Enable post-editing

Generic composite foods can never fully cover the diversity of household recipes, loose goods, and products. Dietary surveys should therefore always allow post-editing of ingredients to reflect actual intake.

Unlike generic single foods, the original form is not “raw” but a mixture of processing states at the ingredient level. In household recipes, ingredients are usually weighed raw. For loose goods, weighing the food in its original form is common. For generic alternatives to branded foods, the weight of the original form is usually known, and further processing often happens at the recipe level (e.g., cooking a pack of pasta).

Missing BLS recipes

Currently, generic composite foods originate from the BLS 4.0. No recipes have been published for this version. To continue supporting mHEI analysis, we introduced LLM-generated recipes.

Development outlook

Due to missing BLS 4.0 recipes, we plan to replace current generic composite foods entirely with our own recipes. This will enable post-editing, improved portion sizes, and processing states at the recipe level. We also consider a new endpoint to recommend household recipes that maximize improvements in evaluation quality.

Branded Foods

Branded foods represent manufacturers’ product ranges (e.g., “Twix”). Unlike generic foods, they capture product-specific variations in ingredients and nutrients (e.g., differences in sugar content between two granolas).

Tip: Show brand for context

Product names alone are often ambiguous (e.g., “Peaches” by Haribo), so brand names should be displayed for context.

Raw data are obtained directly from manufacturers via the GDSN. We only use data that manufacturers publish to the public data pool.

Gaps in GDSN coverage

Additional data sources could be incorporated to improve coverage, but the potential benefit must be weighed against significant additional costs.

Using Large Language Models (LLMs), ingredient lists are parsed and mapped to generic single foods and generic composite foods. Based on the declared nutrients of the branded foods and the nutrients of the generic foods, ingredient amounts are estimated via optimization algorithms and the full nutrient profile is computed. Incomplete or low-quality raw data are excluded during processing.

Limitations of nutrient optimization

Optimization is based on the Big 7 (energy, fat, saturated fat, carbohydrates, sugars, protein, and salt). It is very accurate for macronutrient-rich ingredients but unreliable for small-share components (e.g., vitamin fortifications). This has little impact on mHEI but reduces accuracy for micronutrient estimates.

In contrast to generic composite foods, branded foods reflect actual consumption, so post-editing is unnecessary. Listing all ingredients would be unwieldy anyway due to very long (50+) ingredient lists.

Development outlook

Once processing states at the recipe level are implemented for generic composite foods, we plan to enable them for branded foods as well. We also plan to explicitly account for fermentation (e.g., beer) and higher nutrient variance (e.g., meat) during processing and to further improve data quality.