Associations between the spatiotemporal distribution of Kawasaki disease and environmental factors: evidence supporting a multifactorial etiologic model

Abstract

The etiology of Kawasaki Disease (KD), the most common cause of acquired heart disease in children in developed countries, remains elusive, but could be multifactorial in nature as suggested by the numerous environmental and infectious exposures that have previously been linked to its epidemiology. There is still a lack of a comprehensive model describing these complex associations. We present a Bayesian disease model that provides insight in the spatiotemporal distribution of KD in Canada from 2004 to 2017. The disease model including environmental factors had improved Watanabe-Akaike information criterion (WAIC) compared to the base model which included only spatiotemporal and demographic effects and had excellent performance in recapitulating the spatiotemporal distribution of KD in Canada (98% and 86% spatial and temporal correlations, respectively). The model suggests an association between the distribution of KD and population composition, weather-related factors, aeroallergen exposure, pollution, atmospheric concentration of spores and algae, and the incidence of healthcare encounters for bacterial pneumonia or viral intestinal infections. This model could be the basis of a hypothetical data-driven framework for the spatiotemporal distribution of KD. It also generates novel hypotheses about the etiology of KD, and provides a basis for the future development of a predictive and surveillance model.

Figure 1

Identification of KD patients and theoretical framework of the spatiotemporal matrix for use for modelling. The spatial and temporal scales were held constant for incidence of KD and the distribution of all predictive features so that data points at the same position in the matrix are aligned with each other in time and space. The temporal (1 month) and spatial (group of ~ 10 FSAs) calipers were decided based on density distribution of the incidence of KD. The example below (red bar) shows the alignment of the incidence of KD and of a generic predicting feature at a specific time and location in our spatiotemporal matrix. KD Kawasaki disease, n number, FSA Forward Sortation Area.

Figure 2

Posterior mean value for the temporal and spatial effects for the risk of Kawasaki disease (KD) in Canada from April 2004 to March 2017 based on the Bayesian hierarchical model. Standardized incidence rate above 1 indicate increased risks, and below 1 decreased risks. (a) Monthly temporal unstructured and structured trends. An increasing trend is visible for the unstructured effect (solid line) across time. The structured effect (dashed line) shows nearly no fluctuation. (b) The seasonal effect illustrates a below average standardized incidence rate for KD in September and a high risk in the winter months. (c) Mean spatially autocorrelated random effect, expressed as area-specific standardized incidence rate. Areas with standardized incidence rate larger than 1 have an above average risk compared to the rest of Canada, while standardized incidence rate smaller than 1 indicate a below average risk. The red colours in the Greater Toronto Area and around Halifax for example indicate a 50 to 90% increase of risk. For better visibility of small regions, an interactive map can be found in the online appendix. The spatial units (of the analysis regions consisting of several forward sorting areas) are indicated with boundaries and grey filling. White regions were added to the plot for easier orientation. For uncertainty estimates, see online appendix.

Figure 3

Association of factors with the standardized incidence rate for Kawasaki disease (KD). Distribution of the standardized incidence rate (line = median) based on the posteriors of the fixed effects (same as multivariable results in Table 1). The values represent the change in standardized incidence rates for a 1 standard deviation change in the predictor. A value below 1.0 (blue) indicates a negative association between the predictor and the standardized incidence rate for KD, and values above 1.0 (red) indicate a positive association. The background color of the labels indicates the type of predictor, with grey representing genetic and demographic factors, blue representing atmospheric pollution, red representing the incidence of healthcare encounters for various causes and green representing atmospheric concentration of biological particles. at.c atmospheric concentration, HE healthcare encounters.

Figure 4

Correlation between observed and modelled Kawasaki disease (KD) rates. Modelled rates have been obtained from the exponentiated risk ratios of the multivariable hierarchical Bayesian model posteriors, multiplied by the expected rate. (a) Multi-year mean of observed and modelled KD rates, where each dot represents 1 of the 100 regions (groups of forward sorting areas). (b) Regional mean of observed and modelled KD rates. The rates were aggregated across all of Canada, and each dot represents 1 month of the 13 years. (c) Time series of observed (black) vs modelled (red) KD rates averaged over Canada from April 2004 to March 2017.

Figure 5

Hypothetical framework of Kawasaki disease (KD) etiology and risk based on published evidence, our previously published environmental case control study and the evidence presented in our current spatiotemporal model of the spatiotemporal distribution of KD. Red indicates higher risk while green indicates lower risk.