The Canary Islands are a volcanic archipelago situated in the eastern North Atlantic (figure 1). This study was performed in Tenerife, the largest island (2034.38km2). Meteorology is characterized by trade winds (March to August), which contribute to the rapid dispersion of local air pollutants and help to keep fairly good air quality in most of the archipelago. Although dust events may occur through the year, there are 2 dust seasons,10 segregated by meteorology: a) November to March, and b) July to August.

The study was carried out in a tertiary care hospital (Tenerife, Spain), which provides medical care for the residents in the northern part of the island of Tenerife, serving a population of 343 025 inhabitants. The northern side of Tenerife includes many residential areas, so many of their inhabitants commute each morning to work in other parts of the islands (mostly to the capital, Santa Cruz, and the southern part of Tenerife), and are thus exposed to air quality conditions throughout the island. Our study included all patients diagnosed with ACS, with or without AMI, from December 2012 to December 2017 at that institution. AMI was defined according to current guidelines as the presence of symptoms of myocardial ischemia and an increase in markers of myocardial necrosis.11,12

A number of clinical variables were prospectively recorded in each patient, including coronary risk factors, age, sex, previous vascular disease, presence of asthma or chronic obstructive pulmonary disease, chronic kidney disease, atrial fibrillation, number of diseased coronary vessels (recorded from the coronary angiography), left ventricular ejection fraction, Killip class, troponin I levels, coronary intervention (percutaneous transluminal coronary angioplasty or coronary artery bypass surgery), hospital stay, in-hospital mortality and 1-year mortality.

This study was approved by the Clinical Research Ethics Committee of a tertiary care hospital (Hospital Universitario de Canarias).

Data on meteorology, concentrations of PM10, PM2.5-10, PM2.5 and reactive gasesous pollutants (NOx, O3, and SO2) were prospectivelly collected daily from December 2012 to December 2017. These data were provided by the air quality network of the Canary Islands, which is part of the European Air Quality Network, in which particulate (PM10 and PM2.5) and gaseous pollutants are monitored following the mandatory reference methods and procedures established by the European Union legislation (Directive 2008/50/EC). The network provides hourly resolution data, which are transmitted to the European Environmental Agency. We determined and analyzed daily average values. To obtain a comprehensive view of the impact of Saharan dust events on air quality, we initillay worked with the data of several air quality stations, located in the North (Balsa Zamora, residental background), North-East (Depósito Tristán and Tena Artigas, residential background) and South (El Río, rural) of Tenerife. Because of the rather low background of PM and because Saharan dust events prompt high PM concentrations across the island, the time series of PM at these sites exibited a high correlation. Thus, we selected the records with the highest data availability (Tena Artigas station) for inclusion in the statistical analysis with the medical data. During the study period, the time series of daily PM10 concentrations of Tena Artigas (availability=98.4%) exibited a correlation coeficient within the range 0.76 to 0.84 with the other 3 stations during the study period (December 2012 to December 2017). The occurrence of Saharan dust events was detected with the validated past forecasts of the WMO SDS-WAS (World Meteorological Organization's Sand and Dust Storm – Warning Advisory and Assessment System), whose regional node for Northern Africa, Middle East and Europe is managed by the Meteorological State Agency of Spain and the Barcelona Super Computing Center. The WMO SDS-WAS dust forecast system is constituted by a weather forecast model that also includes a module for simulating the dust (cycle emissions, transport, and deposition).13 This dust forecast for the Canary Islands was validated by García-Castrillo and Terradellas.14

Quantitative and categorical variables are respectively described as median [interquartile range] and number (percentage). We used time-series regression analyses, which are commonly used to study the association between aggregated exposures and health outcomes. The outcome is a count (usually the number of events per day) and the main unit of analysis is the day and not the individual person. Because usual individual confounders (eg, age, sex, diabetes) do not change significantly from day to day, these factors have no impact on effect estimates of environment variables. To control for seasonality and long-term trends, we used a time-stratified case crossover design, with the strata being the day of the week between the month.15,16 The idea was to compare the exposure of a case immediately prior to the event with that same person's exposure at otherwise similar “reference” times. Because our strata was the day of the week within months, exposures associated with an event occurring 1 day of the week, for example, a Monday, were compared with exposures on the rest of the Mondays in the same month and were used as references. These strata were expected to control for individual covariates and slow or regular changes in underlying risk, such as season, trends over time, or day of the week.15,16

A conditional logistic regression model is usually used to analyze the impact of environment variables on clinical events. However, due to the influence of unmeasured causes of clinical events that vary over time, observed event counts usually have wider variation than that predicted by a Poisson distribution.15,16 This overdispersion, when using a logistic regression model, is not taken into account because outcomes are binary and therefore over-dispersion is not apparent. However, the assumption of independence between case-control sets in a conditional logistic model implicitly assumes no overdispersion of counts.16 Moreover, conditional logistic regression assumes that observations are independent. However, observations close in time are probably more similar than those distant in time17 and autocorrelation is likely to be present. In such a scenario, the usual conditional logistic regression model cannot adjust for overdispersion and autocorrelation, both frequent causes of underestimating uncertainty or type 1 error.16

To overcome all these methodological limitations, following the method published by Armstrong et al.,16 we used a conditional Poisson regression model to control for overdispersion and autocorrelation. The dependent variable was the number of ACS diagnosed at the institution per day. To perform this analysis, data matrix had to be modified (figure 1 of the supplementary data). We calculated incidence rate ratios (IRR) with 95% confidence intervals (95%CI). To determine the influence of PM10, we adjusted for the following pollutants and environment variables: PM2.5-10, PM2.5, NO2, SO2, O3, temperature, and humidity. The maximum model was constructed with PM10, the previously described pollutants and environment variables and the 2-way interactions between PM10 and all other variables. All these interactions were assessed and removed if statistical significance was not found. We also took into account the possibility that the impact of the particles was delayed a few days. Thus, we studied the single-day lag effect from day 0 to day 5. We also adjusted these different lag effects for each other, developing a lag-stratified distributed lag model.16 However, in this last model, lag-terms are likely to be highly correlated. To reduce the collinearity of these lag-terms, which results in imprecise estimates, we imposed some constraints to the distributed lag model, namely that effect estimates for days 1 and 2 are the same, and effect estimates for days 3 to 5 are also the same.16 We repeated all analyses, calculating the incidence of AMI instead of ACS. Diagnoses of the model were tested. P values <.05 were considered statistically significant. All analyses were performed using STATA v.15 (StataCorp, United States).

During the study period, 2416 patients were diagnosed with ACS at the aforementioned institution. The median age was 64 (54-74) years and 649 participants (26.86%) were women. A total of 1350 (55.88%) had AMI, 185 (7.67%) were admitted in III or IV Killip class and 129 (5.34%) patients died during the hospital admission. These and other variables are described in table 1.

In the study period, there were 485 days (26.75%) in which there was no ACS, 621 days (34.25%) in which there was 1 ACS, 428 days (23.61%) in which there were 2 ACS, 200 days (11.03%) in which there were 3 ACS, 60 days (3.31%) in which there were 4 ACS, 15 days (0.83%) in which there were 5 ACS, and 4 days (0.22%) in which there were 6 ACS. The median PM10 level was 15μg/m3 (11-24). For those days without ACS, the PM10 concentration was 15μg/m3 (11-23). For days with 1, 2, 3, 4, 5 or 6 ACS, PM10 levels were 16μg/m3 (11-24), 16μg/m3 (11-23), 16μg/m3 (11-24), 16μg/m3 (11-24), and 13μg/m3 (10-23), respectively. Table 2 shows PM10 concentrations for different numbers of ACS/d and different day-lags.

Figure 2A of the supplementary data shows the daily PM10 concentrations in Tenerife during the study period. This time series of PM10 shows the regular low background of this region (∼15 μg/m3) and sharp peak events due to the arrival of dusty air from North Africa, when PM10 reached values between 40 and 350μg/m3. Figure 2B of the supplementary data shows the time series of dust concentrations in Tenerife provided by the WMO SDS-WAS modelling, where the peak dust events (up to ∼320μg/m3) led to the high PM10 episodes, as can be clearly observed.

On the same day of the ACS (with no lag-effect), the IRR for each 10μg/m3 of PM10 was 1.01 (95%CI, 0.80-1.28). The results of the Poisson regression showing IRR of PM10 and other confounding factors are shown in table 3. No long-term trends between the study period and the predicted number of ACS were observed (figure 2).

Figure 2.

Predicted number of ACS/d during the study period based on the Poisson regression model. Line shows a lowest function. ACS, acute coronary syndrome.

(0.11MB).

Study of possible day-lag effects showed that PM10 concentrations were not associated with the number of ACS in any of the previous 5 days. The IRR of PM10 levels of 1, 2, 3, 4 and 5 days (for changes in 10μg/m3) were 1.27 (95%CI, 0.87-1.85), 0.92 (95%CI, 0.84-1.01), 0.74 (95%CI, 0.45-1.22), 0.98 (95%CI, 0.87-1.11), and 0.95 (95%CI, 0.84-1.06), respectively. No association between PM10 and the number of ACS was found when these different lag-effects were adjusted by each other in a model with constraints (effect estimates for day-lag 0, 1 and 2 are the same, and for day-lag 3, 4 and 5 are the same). Figure 3 shows the IRR, which results from modelling lagged associations between PM10 and the number of ACS. Interactions were not statistically significant in any of the models and were eliminated.

Figure 3.

IRR calculated modelling lagged associations between PM10 concentrations and number of ACS. A: without controlling for other lag-effects, B: controlling for different lag-effects with constraints. IRR for changes in 10μg/m3 of PM10. ACS, acute coronary syndrome; IRR, incidence rate ratios; PM10, particulate matter smaller than 10-micron aerodynamic diameter.

(0.14MB).

A total of 1350 patients were diagnosed with AMI. Based on the Poisson regression model adjusted for autocorrelation and overdispersion, PM10 levels had no influence on the number of AMI. For 10μg/m3 of PM10, the IRR was 1.1 (95%CI, 0.71-1.55). The number of AMI was not associated with the PM10 concentrations of any of the previous 5 days. No association between PM10 and the number of ACS was found when these different lag-effects were adjusted by each other in a model with constraints (figure 4).

Figure 4.

IRR calculated modelling lagged associations between PM10 concentrations and number of acute myocardial infarction. A: without controlling for other lag-effects, B: controlling for different lag-effects with constraints. IRR for changes in 10μg/m3 of PM10. IRR, incidence rate ratios; PM10, particulate matter smaller than 10-micron aerodynamic diameter.

(0.15MB).