Multivariate time series clustering analysis of the Global Dietary Database to uncover patterns in dietary trends (1990-2018).

Abstract

Objective: Understanding country-level nutrition intake is crucial to global nutritional policies that aim to reduce disparities and relevant disease burdens. Still, there are limited numbers of studies using clustering techniques to analyse the recent Global Dietary Database (GDD). This study aims to extend an existing multivariate time series (MTS) clustering algorithm to allow for greater customisability and provide the first cluster analysis of the GDD to explore temporal trends in country-level nutrition profiles (1990-2018).

Design: Trends in sugar-sweetened beverage intake and nutritional deficiency were explored using the newly developed programme 'MTSclust'. Time series clustering algorithms are different from simple clustering approaches in their ability to appreciate temporal elements.

Setting: Nutritional and demographical data from 176 countries were analysed from the GDD.

Participants: Population representative samples of the 176 in the GDD.

Results: In a three-class test specific to the domain, the MTSclust programme achieved a mean accuracy of 71·5 % (adjusted Rand Index [ARI] = 0·381) while the mean accuracy of a popular algorithm, DTWclust, was 58 % (ARI = 0·224). The clustering of nutritional deficiency and sugar-sweetened beverage intake identified several common trends among countries and found that these did not change by demographics. MTS clustering demonstrated a global convergence towards a Western diet.

Conclusion: While global nutrition trends are associated with geography, demographic variables such as sex and age are less influential to the trends of certain nutrition intake. The literature could be further supplemented by applying outcome-guided methods to explore how these trends link to disease burdens.