Use case

Food Informatics

Consumers are confronted with an increasingly complex decision-making process when it comes to food choices. The current trend is to provide consumers with more and more data, via printed labels or in digital form. This however only leads to more confusion, uncertainty and indifference. Instead, consumers need personalized and contextualized advice that is attractive and motivating. Validated knowledge should translate individual and contextual data into an advice that indicates which action is preferable. This advice must be given automatically and be based on scientific evidence.

Solution / approach

Although it would be possible to create a separate software solution for each application, we believe that a generic platform can be developed that provides a flexible and modular framework for
all kinds of similar applications. This allows us to adapt to changing
requirements and new insights in nutrition science. Scientifically, the
challenge is to set up a flexible platform that comprises two types of
reasoning. The first uses logics to infer a conclusion from some initial data.
It applies predicate logics for expressing for example food-health or
food-sustainability relations, but also other context relations. It reasons by
adding ‘triples’ in a knowledge base using modular algorithms. Triples are
subject-predicate-object expressions, such as ‘apple – is_a – fruit’, or
‘amount_of_fibre – has_value – 20_gr’, each element of the triple having its own URI (hyperlink). They express data and knowledge in the most flexible way as a connected graph, which is machine readable. The second approach provides estimations for missing or uncertain data and knowledge. We apply Bayesian Belief Networks for dealing with incomplete or inaccurate data, for example to estimate one’s individual food intake from the typical diet of a consumer segment.

The generic and flexible platform forms the basis for the Food Advice Demonstrator. This engine uses proven scientific knowledge about food-health relations (which most current food apps lack), consumer preference data, and food intake data to create personal advice. The first version of the Demonstrator has been realized. The system is now being applied and evaluated within the Personalized Nutrition and Health research program on the use case ‘stimulating a fibre-rich diet’.

(Expected) impact of the approach

This development serves for example software companies that build and deploy consumer apps for personalised advice. Although thousands of food apps already exist, most of them do not support integral decision making and lack proper evidence-based food-health knowledge. App builders can use our advisory engine to design new, innovative applications. The second group of users are food companies that need similar functionality when it comes to food product design. Finally, food researchers who want to evaluate alternative strategies for personalised advice in intervention studies can use the same engine, tuned to their research questions. The results of this project and
knowledge gained are shared through the regular channels of the associated projects and WDCC. It has been used as educational material in the Unilever Learning Program. Several stakeholders other than WUR are involved through the associated project Personalised Nutrition and Health, such as Jumbo, PSinFood, Philips and FrieslandCampina, and also public organisations such as RIVM and Voedingscentrum are stakeholders.

Next steps

The next steps are to develop new ways to extend the incorporated knowledge models using scientific evidence, as for example food-health relations. How to do this in a systematic and efficient way? Secondly, in most cases we are facing multiple criteria for making
decisions, which are often conflicting. For example, how to rank healthy food products based on ‘maximal protein’ and ‘minimal energy’? For this purpose we must be able to establish an optimal filtering and ranking strategy for each application context. These issues are addressed in 2019. In this year we are also extending the demonstrator in order to improve our support for industrial systems developers. For this we will seek collaboration with software companies such as VitalHealth (Philips).

Tools used:

  • Baysian Belief Networks, for example in Netica
  • Java development environment
  • RDF4J triple store

Cooperation with/partners

  • Links with companies through PPS Personal Health & Nutrition

More information

Jan Top

jan.top@wur.nl

+31 317 480 212