Weather and Cardiovascular Diseases in Quebec Using Empirical Mode Decomposition

Climate change has an impact on human health with an increase in mortality associated to extreme meteorological events. Among the diseases influenced by climate change are cardiovascular diseases (CVD) which represent the second cause of mortality in Quebec. Therefore, the understanding and quantification of the influence of weather on CVD is an important issue for future years. The difficulty to quantify the influence of weather on health lies partly in the complexity of their cycles. Indeed, weather processes are often nonstationary, interdependent and can present several levels of influence which can make the true influence difficult to detect. To overcome such drawbacks, empirical mode decomposition (EMD) is performed on weather data series to obtain components called intrinsic mode functions (IMF). An IMF represents one mode of oscillation of the original series, e.g. annual variations of temperatures. They are basic oscillations allowing little changes in amplitude and frequency over the IMF. They are (quasi) stationary and orthogonal to each other. Moreover, the sum of all the obtained IMFs from a series is equal to this series, meaning that there is no loss of information. The methodology proposed here consists in using IMFs obtained from weather series as new covariates in regression models to explain health issues, applied to CVDs in Quebec (Canada). This approach allows to explain CVDs according to the different modes of variations of the weather, e.g. the influence of weekly variations alone. Since each weather series is decomposed in several IMFs (at least ten in our case), the whole number of new covariates increases quickly. The Lasso method is able to manage a large number of covariates to shrink the regression model and keep only few covariates. This method allows to outline the most important IMFs on the triggering of CVDs. For instance, this method outlines the importance of annual variation of temperatures and long term trends of temperatures and precipitation on CVD deaths in Montreal. Furthermore, EMD also allows decomposing mortality and morbidity CVD series and separates their different scales of variations. Then, undesired variations, such as the weekend effect on hospitalizations, can be removed. Moreover, each remaining IMF of CVD series can be separately introduced as new response variables in regression models with weather IMFs as covariates. This could help to outline the effect of different weather variables at different scales. Preliminary results of the Lasso in a linear fashion show that, in Montreal, ten days lagged annual variations and long term trend of temperatures along with the trend of precipitations have a major influence on CVD mortality. Those three components explain the annual seasonality and the trend of deaths.