Global Environment Monitoring System - Food Contamination Monitoring and Assessment Programme (GEMS/Food)
GEMS/Food Cluster Diets
Originally developed by the World Health Organization (WHO) to predict dietary exposure to radionuclides in food following the Chernobyl accident, the first Regional Diets were issued in 1989. They were derived from FAO Food Balance Sheets to represent five regional dietary patterns, namely Middle Eastern, Far Eastern, African, Latin American and European. A few years later, the Regional Diets were used to predict the exposure to various chemicals occurring in food (e.g. pesticide residues). For this purpose, the need for regional grouping becomes less important than the one for a grouping based on similarities between the diets.
In 1997, WHO introduced the GEMS/Food cluster diets, as part of the mandate of the Global Environment Monitoring System - Food Contamination Monitoring and Assessment Programme, which is commonly known as GEMS/Food. First cluster diets were initially based on the 1990-1994 FAO food supply data. The method used cluster analysis and an iterative approach based on the use of nineteen marker foods chosen from the foods reported and geographic locations, to define 13 diets representing a total of 183 countries. The 13 cluster diets, including 383 different food items, were later updated using food supply data from 1997 to 2001.
GEMS/Food cluster diets 2012
The Cluster Diets should be used when the contamination is likely to be similar worldwide. In that case the type of food consumed is predominant over the region of consumption. An example is the dietary exposure to pesticide residues.
In 2012 the cluster diets* are based on a newly available methodology for clustering which can be applied on large matrices including a lot of null values. The modelling consists at first in extracting “latent variables” called Consumption Systems (CS) that generate a substructure from the initial database in which countries’ food consumption is expressed and then by deriving a clustering of countries according to their consumption profiles defined by the CS. Such data processing is possible by the utilization of two recently developed statistical learning methods: Non-negative Matrix Factorisation (NMF) and Sparse k-Means Clustering (SKMC). The estimated level of per capita consumption is the average of food supply data for the period ranging between 2002 and 2007. SUA data were processed without any correction, such as under- or over-reporting corrections, only missing data or not reported values were taken as zero for computational reasons. All data were represented as a matrix (with real non-negative entries) corresponding to the estimated per capita consumption with 415 lines of primary or semi-processed food products, and 179 columns, one for each country for which a SUA data was available. Finally data were grouped into 62 food groups.
*New approach for the assessment of cluster diets. Sy M. M., Feinberg M., Verger P., Barré T., Clémençon S. and Crépet A. Food and Chemical Toxicology, 2012, In press.