Health impacts of wildfire-related air pollution in Brazil: a nationwide study of more than 2 million hospital admissions between 2008 and 2018

exposure data

Wildfire data

We accessed the forest fire data from the National Institute for Spatial Research in Brazil – Instituto Nacional de Pesquisas Espaciais – INPE (http://queimadas.dgi.inpe.br/queimadas/). The data obtained contains wildfire records, including the date of occurrence of the forest fire and the geographical location. These data are from seven remote sensing satellite observations, including NOAA-18, NOAA-19, METOP-B, and the Medium Resolution Imaging Spectroradiometer (MODIS) (NASA TERRA and AQUA). ), VIIRS (NPP-Suomi and NOAA-20), GOES-16, and MSG-3. INPE processes all images from these satellites and then estimates the occurrence of forest fires using a specific satellite as a reference. Currently, Aqua is the reference satellite. We calculated all wildfire records in Brazil based on the reference satellite between 2008 and 2018. Since everyone in the health data is based on the municipality level (details of health data are available in the “Health and Population Data” section), we use GIS technologies To summarize the number of wildfire incidents in each Brazilian municipality.There are 5,572 municipalities in Brazil, which represent the smallest areas considered by the Brazilian political system.The government divides the municipalities into five regions, including North, Northeast, Midwest, Southeast and South.Supplementary Figure 1 , we show the spatial distribution of all municipalities and regions in Brazil.

We defined the concept of ‘forest fire wave’ as any day on which wildfires and PM . are recorded_2.5 Concentration exceeded the 99th percentile of the time series from 2008 to 2018 by Brazilian region (forest fires and particulate matter_2.5 data, respectively). We used this concept to capture periods of high incidence of wildfires, allowing us to estimate the health effects associated with strong episodes of bushfire-related air pollution. The concept of a wildfire wave identified in our study is similar to the concepts of severe air pollution events from wildfires identified in previous studies^{11, 15}.

air pollution data

Air pollution data were obtained from cluster models. We accounted for daily PM_2.5 (µg m⁻³), carbon monoxide (ppb), no₂ (ppb) and O₃ (ppb) Concentrations from 2008 to 2018. Data were accessed from the Health Environmental Information System (http://queimadas.dgi.inpe.br/queimadas/sisam/v2/dados/download/). This is a database system developed by INPE – the National Institute for Spatial Research in Brazil.

INPE obtained daily concentrations of all four of these contaminants from ECMWF. ECMWF manages services related to meteorology and air pollution covariates, and implements the Copernicus Atmospheric Monitoring Service (CAMS) on behalf of the European Union including CAMS-Reanalysis and CAMS Near Real-time (CAMS-NRT) forecasts. The CAS service runs ensemble models using various satellite observations and emission inventories among other forecasters. We acquired the data with a spatial resolution of 0.125° (~12.5 km) and a temporal resolution of 6 hours, including daily estimates of 00, 06, 12 and 18 UTC. In our analyses, we used CAMS-Reanalysis for the period between 2008 and 2017, and CAMS-NRT for 2018.

The validation of the global model for a competency assurance management system was reported by Inness et al.³⁶. Specifically, for PM_2.5, the exposure variable in our study, is assessed by ground observations of the Aerosol Robotic Network (AERONET). There are more than 500 AERONET stations worldwide that measure spectroscopic aerosol optical depth (AOD) using terrestrial sun photometers. Among those AERONET stations, about 27 are in Brazil. Validation by Inness et al.³⁶. Estimate the mean biases and SDs from the data provided by the satellite instruments (included in the CAMS aerosol model) relative to the AERONET data. In South America, data from satellite instruments are slightly smaller, with an approximate bias of −0.006 ± 0.128. Another investigation shows that CAM estimates in South America have an rms error (compared to AERONET stations) of 0.268.³⁷. Other studies have shown that AERONET observation sites in South America have a significant representation of AOD measured by MODIS, aboard the TERRA and AQUA satellites.³⁸. It is noteworthy that MODIS is a tool included in the CAMS model. This correlation between AERONET and AOD data from MODIS is significant during biomass burning seasons in South America, which NS² (Coefficient of Determination) for most AERONET stations in Brazil was higher than 0.85³⁸.

We calculated the average daily temporal resolution for each pollutant. Finally, we compiled air pollution data by the municipality, taking into account the geographical location of the headquarters of each municipality in Brazil.

As mentioned earlier, in this study, exposure defined as air pollution associated with wildfires was based on fine particulate matter._2.5 Concentration, when it exceeded the 99th percentile of the time series. This cut-off point is close to the World Health Organization’s 24-hour air quality standard for fine particulates_2.5 (25 micrograms m⁻³). Using this 99% threshold, our findings allow it to be useful for public health agencies to act when air pollution standards are higher. As described in the Introduction section, the literature has reported PM_2.5 As the main pollutant emitted by forest fires. Other pollutants were included in our analyzes as control variables. In the Statistical Analyses section, we describe the statistical model with all control and confounding variables.

weather data

Meteorological data from ensemble models were also retrieved, and accessed from the ECMWF. Weather data includes surface temperature (°C), humidity (%), wind speed (m/s), wind direction (°), and precipitation (mm/day). Temperature, humidity, wind speed and wind direction were derived from Era-Interim reanalysis, with a spatial resolution of 0.125° and a temporal resolution of 6 h. This analysis was performed by ECMWF. Precipitation data was accessed from the Climate Prediction Center and NOAA. These data have an original spatial resolution of 0.50° (~50 km), with an interpolation of 12.5 km, and a temporal resolution of 6 h. For the air pollution data, we calculated the daily average time resolution for each weather variable and then grouped the data by municipality.

health and population data

Hospital admission data was provided by the Ministry of Health of Brazil. These data were obtained from the publicly available database (https://bigdata-metadados.icct.fiocruz.br/dataset/sistema-de-informacoes-hospitalares-do-sus-sihsus) sponsored by the Ministry of Health of Brazil. The data include individual records of hospital admissions in Brazil between 2008 and 2018. Analysis of this data was approved by the database management group.

Hospital admission information included event date, municipality of home, age, gender, ethnicity, number of days patients spent in hospital, and main diagnosis according to International Classification of Diseases, version 10 (ICD-10) codes. As discussed in the Introduction, review studies of the health effects of wildfire exposure report consistent evidence for an association between wildfire exposure and cardiorespiratory health effects.^10,39. Therefore, in this study we examined respiratory disease (ICD-10 codes J00-J99) and cardiovascular disease (ICD-10 codes I00-I99). During the period between 2008 and 2018, there were 2,044,038 hospitalizations for cardio-respiratory disease in Brazil.

Statistical analysis

We applied a time stratified cross-case study design using conditional logistic regression models. The design of this study is based on a binary indicator variable of case/observation days to compare exposure (PM) associated with forest fires_2.5 in bushfire waves) on the day of the health event (hospitalization; case day) with exposure on days when the event did not occur (observation days). We used time-layer sampling to determine reference exposure days, which were matched to the day of the week, month and calendar year of hospital admission. This allows comparing exposure on the day of the health event on a Monday in January of 2008, for example, with exposure on all other Mondays in January of 2008. It should be noted that in the design of this study, each case period includes three or four control periods.

We chose to perform a conformational analysis, because exposure to air pollution associated with wildfires is an accidental event. Moreover, the matching approach has some advantages. First, because the matching periods were close in time, the approach minimizes the effects of confounding associated with seasonal trend by controlling for time-dependent risk factors, including day of the week, season, and long-term trends by matching. Also, subjects exposed to the health event were defined as their own controls, allowing all potential confounders to be controlled at the individual level (eg, socioeconomic status, smoking history, pre-existing medical conditions) by design, except where change quickly.

We used a conditional logistic regression model to estimate the odds ratio (OR) for hospital admission associated with wildfire-related particles._2.5 In wildfire waves compared to the background. We modified the model for several control/confounding variables, including other air pollutants emitted by wildfires (CO, NO).₂, and O₃), meteorological variables (temperature, relative humidity, precipitation, wind speed, wind direction), topographical variable represented by altitude, spatial terms (latitude, longitude, condition, binary variable representing municipalities representing capitals), health variable It indicates the number of days patients spend in the hospital.

In the primary analysis, we applied the conditional logistic model described above for each group of health outcomes—respiratory disease and cardiovascular disease. We considered moving averages of wildfires, air pollutants, and weather variables. We considered five moving averages, including the one-day moving average, the two-day moving average, …, the 5-day moving average.

We performed several effect adjustments and sensitivity analyzes by stratifying analyzes by sex, by age (0–5 years, 35–64 years, >64 years), by excluding control pollutants from the model (CO, NO₂, and O₃), by excluding race, excluding some spatial terms (state and latitude/longitude), and counting cardiovascular and respiratory hospital admissions together. We applied a moving average to each of these layers (subgroup analysis). All statistical analyzes were performed in R, using the Survival statistical package (clogit function).

Primary analysis and all sensitivity analyzes were performed individually for each of the five Brazilian regions (Appendix 1 shows the spatial distribution of these regions). We performed this subgroup analysis by region to capture the regional variability of Brazil’s landscapes (eg, Amazonian forests, Atlantic forests, Pantanal, etc.), which is closely related to the occurrence of wildfires. Next, the region-specific OR was meta-analyzed to estimate the national median hospital admission associated with wildfire-related particles._2.5 In waves of forest fires. We accounted for intra- and inter-area variance by applying a regression meta-analysis with random effects.

Report summary

More information about research design is available in the Nature Research Reporting Summary linked to this article.