Estimated dietary pesticide exposure from plant-based foods using NMF-derived profiles in a large sample of French adults

Abstract

Purpose

This study, conducted in participants of the NutriNet-Santé cohort, aims to identify dietary pesticide exposure profiles (derived from Non-negative Matrix Factorization) from conventional and organic foods among a large sample of general population French adults.

Methods

Organic and conventional dietary intakes were assessed using a self-administered semi-quantitative food frequency questionnaire. Exposure to 25 commonly used pesticides was evaluated using food contamination data from Chemisches und Veterinäruntersuchungsamt Stuttgart accounting for farming system (organic or conventional). Dietary pesticide exposure profiles were identified using Non-Negative Matrix factorization (NMF), especially adapted for non-negative data with excess zeros. The NMF scores were introduced in a hierarchical clustering process.

Results

Overall, the identified clusters (N = 34,193) seemed to be exposed to the same compounds with gradual intensity. Cluster 1 displayed the lowest energy intake and estimated dietary pesticide exposure, high organic food (OF) consumption (23.3%) and a higher proportion of male participants than other groups. Clusters 2 and 5 presented intermediate energy intake, lower OF consumption and intermediate estimated pesticide exposure. Cluster 3 showed high conventional fruits and vegetable (FV) intake, high estimated pesticide exposure, and fewer smokers. Cluster 4 estimated pesticide exposure varied more across compounds than for other clusters, with highest estimated exposures for acetamiprid, azadirachtin, cypermethrin, pyrethrins, spinosad. OF proportion in the diet was the highest (31.5%).

Conclusion

Estimated dietary pesticide exposures appeared to vary across the clusters and to be related to OF proportion in the diet.

Trial registration

Clinical Trial Registry: NCT03335644.

Appendices

Appendix 1

Detailed explanation for the non-negative matrix factorization procedure

Non-negative Matrix Factorization is a non-supervised data decomposition method, proposed by Lee to deal with non-negative data using non-negativity constraints.

This method is relevant for non-negative data with excess zeros and measurement error such as exposure to pesticides constrained by the detection limits of dosing techniques.

The purpose of NMF is to explain observed data through a limited number of components approximating the original data as accurately as possible.

The matrix representing the basis components and the matrix of mixture coefficients are constrained to have non-negative values, and no orthogonality or independence constraints are imposed on the basis components.

Let X be a matrix (n × p) containing only non-negative values and without a row or column containing only 0 and r a relatively small integer < n and < p.

The non-negative factorization of matrix X is the search for two matrices W (n × r) and H (r × p) containing only positive or zero values and whose product approaches X so that X ≈ WH.

The factorization is solved by searching for a local optimum of the optimization problem:

$$\mathrm{min} \,W, H \geq 0 [L (X, WH)].$$

L is a loss function measuring approximation quality. Since the objective is usually to reduce the dimension of the original data, the factorization rank r is in practice often chosen such that r < min(n, p). This equation is solved by a multiplicative algorithm based on a gradient descent approach.

In this study of dietary pesticide profiles, W would be the total dietary exposure to the 25 selected pesticides (previously obtained after combining contamination values for each food and foods consumed by each participant) and H the number of individuals

K, the number of NMF Components.

(Adapted from Zetlaoui et al. 2011).

Appendix 2

Flowchart for the decomposition of ingredients and matching

Selected Ingredients included in the pesticides database: