Nutrition labels disagree across rating systems
Front-of-pack nutrition labels were built independently by different national bodies, on different evidence bases, for different food environments. The same yoghurt can read as a green-light food on one system, a mid-grade letter on another, and an ultra-processed item on a third. For someone trying to compare two foods on a shelf, the labels do not agree, and most consumers see only one of them at a time.
The problem compounds when the food is not a Western retail package. NutriScan is a classifier that brings three already-public rating systems together over a database that includes the foods existing tools tend to miss.
- Labels disagree at the food level. The same item can read as healthy on one published scheme and as ultra-processed on another, and consumers usually see only one reading at a time.
- Coverage is biased toward Western retail. Existing rating tools were built around supermarket panels and recognizable brand names, and they fall back to guesses on foods that do not arrive in a box.
- South Asian and home-cooked foods are systematically under-represented. Regional staples, household dishes, and foods that move through the supply chain without a barcode rarely receive a consistent rating.
- Integration, not invention. The methods are all already published. The contribution is to apply them together over a broader database than any single tool currently covers.
Public food composition databases
The classifier draws from public food composition databases that publish nutrient-per-100g values for raw and prepared foods. The reference set spans Western retail packaging data, US public nutrient data, and Indian food composition tables, with global open-product coverage as a fallback. Specific dataset identifiers and the curated fallback set are held until publication.
Each food is looked up against the available sources, with conflicts resolved through a fixed precedence and a curated fallback database for foods that are not well-covered by any single source.
Three published rating systems, applied in parallel
NutriScan does not introduce a new scoring method. It combines three peer-reviewed and publicly documented ones:
- Nutri-Score (A to E). Front-of-pack nutrition grade developed by Santé publique France. A is best, E is worst. The algorithm is published.
- UK FSA traffic-light indicators. Red, amber, and green markers per serving for fat, saturated fat, sugar, and salt. The thresholds are published by the UK Food Standards Agency.
- NOVA processing level (1 to 4). Classification of foods by degree of industrial processing, originally developed at the University of São Paulo.
For each food, NutriScan emits all three readings side by side. Where the three disagree, the user sees the disagreement rather than a forced consensus. The classifier's contribution is integration and coverage, not a new judgment about what a food is.
One classifier for packaged, Indian and home-cooked foods
Coverage of Indian and home-cooked foods is a deliberate scope choice. Most existing nutrition-rating tools were built around Western retail packaging, and they fall back to "unknown" or guess wrong on dishes that do not arrive in a box. NutriScan is built to close that gap by combining multiple food composition sources and a curated fallback set for common dishes that are otherwise poorly covered.
The tool is designed to be used at the point of decision: search a food by name, scan a barcode where one exists, and see a composite reading that names which system said what. Where a food is not covered by any source, the tool says so rather than guessing.
Who it is for
A multi-source view of nutritional quality is useful across three audiences: public-health researchers, nutrition educators working in mixed-cuisine populations, and consumers comparing foods at the point of decision.
- Who benefits. Researchers gain a food-level read on where national labeling schemes disagree; educators gain a tool that does not silently fail on home-cooked dishes; consumers see three independent readings on a single screen.
- What is open. The three underlying methods are all already published. The integration concept, the coverage scope, and a public interface for side-by-side comparison are intended for public release.
- What is held. Internal thresholds, the conflict-resolution rules, the curated fallback set, and the validation methodology are held until publication.
- What comes next. Research is in preparation. The classifier and its accompanying methods note will be released after publication.
Current status
In preparation. A working prototype has been built and the research is being written up. Public release will follow publication.