Burden of disease attributed to ambient air pollution in Thailand: A GIS-based approach

Background Growing urbanisation and population requiring enhanced electricity generation as well as the increasing numbers of fossil fuel in Thailand pose important challenges to air quality management which impacts on the health of the population. Mortality attributed to ambient air pollution is one of the sustainable development goals (SDGs). We estimated the spatial pattern of mortality burden attributable to selected ambient air pollution in 2009 based on the empirical evidence in Thailand. Methods We estimated the burden of disease attributable to ambient air pollution based on the comparative risk assessment (CRA) framework developed by the World Health Organization (WHO) and the Global Burden of Disease study (GBD). We integrated geographical information systems (GIS)-based exposure assessments into spatial interpolation models to estimate ambient air pollutant concentrations, the population distribution of exposure and the concentration-response (CR) relationship to quantify ambient air pollution exposure and associated mortality. We obtained air quality data from the Pollution Control Department (PCD) of Thailand surface air pollution monitoring network sources and estimated the CR relationship between relative risk (RR) and concentration of air pollutants from the epidemiological literature. Results We estimated 650–38,410 ambient air pollution-related fatalities and 160–5,982 fatalities that could have been avoided with a 20 reduction in ambient air pollutant concentrations. The summation of population-attributable fraction (PAF) of the disease burden for all-causes mortality in adults due to NO2 and PM2.5 were the highest among all air pollutants at 10% and 7.5%, respectively. The PAF summation of PM2.5 for lung cancer and cardiovascular disease were 16.8% and 14.6% respectively and the PAF summations of mortality attributable to PM10 was 3.4% for all-causes mortality, 1.7% for respiratory and 3.8% for cardiovascular mortality, while the PAF summation of mortality attributable to NO2 was 7.8% for respiratory mortality in Thailand. Conclusion Mortality due to ambient air pollution in Thailand varies across the country. Geographical distribution estimates can identify high exposure areas for planners and policy-makers. Our results suggest that the benefits of a 20% reduction in ambient air pollution concentration could prevent up to 25% of avoidable fatalities each year in all-causes, respiratory and cardiovascular categories. Furthermore, our findings can provide guidelines for future epidemiological investigations and policy decisions to achieve the SDGs.


Introduction
observed data from air monitoring measurements. We utilised GIS to explore the spatial variability of air pollution exposure, and adopted the CRA method developed by the WHO, the GBD [18] and others [19,20] to quantify mortality attributable to ambient air pollution.

Methods
Overall approach to estimate the burden attributable to ambient air pollution We employed the CRA framework which is defined as the systematic evaluation of the changes in the population health and ranking the different factors that contribute to the specific outcome to quantify the burden of disease attributable to ambient air pollution [21]. The general framework and its components are presented in Fig 1. Each component of the estimation is described as follows: • Ambient air pollution exposure • Population distribution of exposure (Pe) • Relative risks and concentration-response relationships • Attributable mortality due to ambient air pollution Many studies have indicated that the health effects of PM 2.5 are more harmful than PM 10 [23][24][25][26][27] with their concentrations highly correlated [28]. The PM 10 correlation was also very high with NO 2 but low with SO 2 [29]. PM 10 was reported by all existing air quality network in Thailand whereas only a few stations reported PM 2.5 concentrations, making analysis at the country level very difficult. Since PM 2.5 is a component of PM 10 , it is possible to estimate PM 2.5 from PM 10 data based on the typical relationships between pollutants, as PM 2.5 can be treated as a fixed weight fraction of PM 10 . We decided to convert PM 10 to PM 2.5 for stations without PM 2.5 readings using the ratio of PM 2.5 to PM 10 based on the literature review. Many studies [30][31][32] have reported PM 2.5 and PM 10 ratios in the range of 0.35-0.7, and a local study in Thailand reported ratios of about 0.5 in Bangkok [33] and clearly stated that daily PM 2.5 and PM 10 concentrations were highly correlated (r ! 0.85) [33]. From the literature review, we decided to use the ratios for PM 2.5 and PM 10 from the local study in Thailand[33] which was similar to the WHO global analysis of disease burden due to outdoor air pollution in developing countries [19].
To estimate the exposure level of air pollutants across geographical area, we used inverse distance weighted (IDW) interpolation method [34][35][36][37][38][39] to estimate the spatiotemporal distribution of ambient air concentrations based on empirical data from the air quality monitoring stations across Thailand and a grid consisting of 40767 cells (3×3 km 2 resolution). For crossvalidation evaluation, comparisons of predicted values to observed values were essential information about the quality of the model [40,41] using all existing data to estimate the trend and model autocorrelation. This removed one or more data locations and predicted their associated data using information from the other locations. We then assessed the accuracy of the model using the square root of the mean for the squared prediction errors (RMSE) based on the predicted and actual values at the existing point. In addition, we obtained cross-validation correlations using the squared Pearson correlation between the measured values at knownpoint observations and the spatial model predictions.
To quantify the different levels of exposure to ambient air pollution for estimation of avoidable disease burden, we used a 20% reduction in ambient air pollutant as a reasonable proportion of reduction in air pollutant level from the Mexico City Air Quality Management Team [42] for possible suggestion scenarios.

Population distribution of exposure (Pe)
To quantify the Pe, we acquired population data from the 2000 Gridded Population of the World, Version 3 (GPWv3), generated by the SEDAC (Socioeconomic Data and Applications Center) project at Columbia University [43]. This dataset was estimated from the human population from national and subnational input sources (usually administrative sources) of varying resolutions into regular latitude-longitude grids at a resolution of 2.5 arc-minute grid cells (or~5 km at the equator). Pe was estimated as the proportion of the population by selected age groups counted in the grid divided by the total population in Thailand. We used population fractions from the Department of Provincial Administration, Thailand [44]. The total population of Thailand in 2009 was about 63 million, spread over an area of 514,000 square kilometres. We assumed the proportions of population located within each grid to have been exposed to the same pollutant concentrations in the grid cells. Pe value are presented in Fig 2. The minimum Pe per grid was zero, while the maximum was 6.9 × 10 −3 . The Pe(s) for grids in Bangkok and the vicinity were relatively high compared to other regions defined in Fig 2, reflecting higher population densities.

Relative risks and concentration-response relationships
Air pollutants such as PM 10 , PM 2.5 , and NO 2 can cause a variety of detrimental public health effects including cardiovascular disease (CVD), respiratory disease, and lung cancer [3,33,45,46]. Relative risk (RR) is commonly used to represent the results of exposure-response functions. This study estimated health impact associated with ambient air pollution using exposure to the risk of mortality based on the relationship between RR, concentration-response coefficient and ambient air pollution concentrations [19][20][21]47]. The health impact function was defined as follows: where β is the concentration-response coefficient (CR), as the slope of the log-linear relationship between ambient air pollution concentrations and mortality, and x-x 0 or ΔX is the concentration change from baseline conditions or natural background concentration. We assumed natural background concentrations of 10 μgm -3 and 3 μgm -3 for PM 10 and PM 2.5 , respectively, based on the WHO environmental burden of disease (EBD) study [19]. For NO 2 , we assumed no background concentrations (zero concentrations) for Thailand. Table 1 summarises estimations on the RRs based on the health impact function. We selected RR for short-and long-term effects based on available local study, systematic reviews and meta-analysis, as well as recommendations from previous EBD studies of ambient air pollutants. The chosen health outcomes were grouped following ICD-10 classification for all-causes mortality, except for deaths attributed to external causes (ICD-10: V01-Y89), cardiovascular disease (ICD-10: I00-I99), and respiratory disease (ICD-10: J00-J99). For PM 10 , we selected RR for all-causes mortality, respiratory and cardiovascular outcome based on available local study in Thailand [9]. The health outcomes of PM 2.5 included all-causes mortality, cardiopulmonary mortality and lung cancer mortality in the population aged > 30 years, all due to long-term exposure to PM 2.5 , using annual average concentration as the exposure indicator [3]. Due to a lack of RR information in Thailand, this study used RR information for NO 2 based on the systematic review and meta-analysis of 23 long-term studies on a global scale, published from 2004 to 2013, evaluating the relationship between NO 2 and mortality outcome [48].

Attributable mortality due to ambient air pollution
To investigate the magnitude of the disease burden attributable to ambient air pollution and mortality associated with ambient air pollution, exposure is expressed as the fraction of disease or death attributable to the risk factor in a population and referred to as the population-attributable fraction (PAF) [18,21]. The PAF has long been used to estimate the proportion reduction of burden that can be attributed to specified risk factors [51,52]. The exposed population may be divided into multiple categories based on the level or length of exposure, each with its own RR. With multiple (n) exposure categories, the PAF is given by the following generalised equation [18]: PAF = proportion of disease burden attributable to ambient air pollution Pe i = proportion estimates of the population in exposure category i, including the unexposed RR i = relative risk (magnitude of the association between ambient air pollution and disease) in exposure category "i", compared to the reference level We calculated PAF using Eq 3 and performed GIS raster algebra analysis using the raster resampling technique for different resolutions. The coarse grid (Pe) served as the basis for our estimate of the total burden of disease across Thailand. To calculate the expected number of mortality cases due to ambient air pollution exposure (E), we applied PAF to the number of mortalities as the following equation; E = expected number of deaths due to ambient air pollution N = baseline number of deaths for each disease outcome The number of disease specific deaths was obtained from the Thai Burden of Disease (BOD) study [53] conducted every five year to provide burden of disease information to setting national health planning priorities. The BOD study estimated age-, sex-, and cause-specific mortality by verifying cause of death from the national vital registration with a nationwide verbal autopsy (VA) study [53][54][55]. The VA study was conducted in 2005 based on a sample of 3,316 in-hospital and 6,328 outside-hospital deaths from 28 selected districts in nine provinces [56]. Completeness adjustment of the vital registration was based on the mid census Survey of Population Change (SPC) conducted by the National Statistical Office [57,58]. Table 2 shows the statistics of average change in concentration (ΔX) values and statistical summary of model performance (best fit) corresponding to a spatial interpolation model for PM 10 , PM 2.5 , and NO 2 from spatial interpolation based on surface monitoring measurements across Thailand.  (Table 2).

Ambient air pollution concentrations in Thailand
We calculated coefficients of correlation between the model predictions and the measured values at the monitoring stations. The Pearson correlation coefficients between best-fit models and actual concentrations for PM 10 , PM 2.5 and NO 2 were 0.44 (95%CI: 0.2-0.7), 0.75 (95%CI: 0.7-0.8), and 0.63 (95%CI: 0.5-0.75), respectively, which were statistically significant for measured ambient air concentrations (p-value <0.01). Fig 4 visualises the geographical distribution of annual mean estimate for exposure to ambient air pollutants across the study period. The exposure estimated for PM 10 , PM 2.5 , NO 2 concentrations appeared in a range from 16.9 μgm -3 to 84.1 μgm -3 , 10.4 μgm -3 to 44.1 μgm -3 and 3 ppb to 48 ppb, respectively, indicating that the Bangkok Metropolitan Region was more polluted than other regions in Thailand.

Health impact and the population attributable fraction (PAF)
The health impact of ambient air pollutants was based on the relationship between change in concentrations (ΔX) and RR as described in Eq 1. Table 2 presented the estimation of RR in each pollutant and health outcome. The spatial distribution of RR across the country indicated a range from 1.01 to 1.35, depending on air pollutants and health outcome. Subsequently, this study determined the summation of PAF grids based on Eq 2 for each pollutant in Thailand in 2009, as shown in Table 3. This study estimated the average PAF grid for all-causes mortality due to PM 10 , PM 2.5 , NO 2 at approximately 1.41 x 10 −6 (95% CI: 1.35 x 10 −6 -1.47 x 10 −6 ), 3.31 x 10 −6 (95% CI: 3.17 x 10 −6 -3.45 x 10 −6 ), and 4.6 x 10 −6 (95% CI: 4.2 x 10 −7 -4.8 x 10 −6 ), respectively. The average PAF for lung cancer caused by PM 2.5 was approximately 8.16 x 10 −6 (95% CI: 7.8 x 10 −6 -8.5 x 10 −6 ). Fig 5 illustrates that the spatial variability of PAF due to long-term ambient air pollution exposure varied across Thailand. The results of this study indicated that the Bangkok Metropolitan Area had the largest percentage of total mortality attributable to PM 2.5 across all ages (level ranged widely from 2.93 x 10 −4 to 7.4 x 10 −4 depending on the risk estimate used), which was the highest among the three air pollutants. The largest percentage of mortality attributable  to NO 2 was also the highest in the Bangkok Metropolitan Area (PAF ranged between 1.67 x 10 −4 to 4.51 x 10 −4 depending on the selected health end-point).
The summation of the PAFs for all grids in the category of air pollutants and disease outcomes based on Eq 2 and Table 1 are represented in Table 3. The PAFs for all-causes mortality for PM 10 , PM 2.5 , and NO 2 were approximately 0.02, 0.1 and 0.1, respectively. The PAFs for respiratory mortality caused by PM 10 and NO 2 were approximately 0.02 and 0.07, respectively; PAFs for cardiovascular mortality caused by PM 10 and PM 2.5 were approximately 0.04 and 0.15, respectively, while PAF for lung cancer caused by PM 2.5 was approximately 0.17. PM 2.5 had the highest model estimated PAF at 17% of lung cancer burden. Table 4 indicates the results for mortality caused by ambient air pollutants. The best estimate demonstrated ambient air pollution-related mortality, which included pollution from PM 2.5 , PM 10 and NO 2 . Annually, there are about 3652-38410, 653-934 and 4024-15361 cases of all-causes, respiratory and cardiovascular mortality, respectively, from long-term exposure to PM and NO 2 , reflecting the highest CR from PM and the underlying cause-specific mortality for each health outcome.
The results in this study indicated that, if PM and NO 2 were reduced by 20% from current levels, the health burden could be reduced in about 5982 cases for all-causes mortality, 160-581 and 146-3401 cases for respiratory and cardiovascular mortality, respectively, depending on each pollutant in Thailand. Similarly, the health burden would have been reduced annually to about 3081 cases for all-causes mortality per year if the highest concentrations for ambient particulate matter (PM 10 ) across the country, which was about 84.6 μg/m 3 , had been reduced to 65 μg/m 3 . Respiratory mortality attributed to PM 10 could be reduced annually to about 160 cases per year, or about 24.4% of the current estimate due to respiratory mortality. Other health outcomes of particulate matter, such as cardiovascular and lung cancer, could also be reduced annually to about 146 and 542 cases per year, respectively.

Discussion
We presented a combination of GIS spatial analysis and empirical information on CRA to quantify the geographical distribution of PAFs and 2009 mortality due to various ambient air pollutants (PM 2.5 , PM 10 and NO 2 ) across Thailand, based on available empirical data. We predicted mortality attributable to short-and long-term ambient air pollution exposure ranging between 933 and 27 thousands persons depending on air pollutants across the country. PAFs varied across the country, as expected; PAF and exposure to air pollutants were relatively concentrated in the Bangkok metropolitan area, which had the largest number of monitoring stations, population density and air pollutant concentrations in Thailand, especially for PM and NO 2 .
Our estimate of all causes mortality attributable to PM 2.5 , was about 38,410 deaths or 6% of total deaths in Thailand. This proportion was not much different in terms of proportion compared to the GBD 2015 [5], which had estimated 7.6% of total mortality for long-term exposure to PM 2.5 globally. GBD used existing surface monitoring data to assemble a georeferenced global PM 2.5 measurement database of 2005 annual average concentrations from available national/regional/local air quality monitoring reports and excluded PM 10 and NO 2 from their estimation [59]. The surface monitoring measurements dataset for the Asia region was based primarily on measured PM 2.5 and appeared in the annual ambient air quality monitoring report from Australia and New Zealand [60]. The surface monitoring measurement datasets for other Asian countries (e.g. Thailand) were obtained from the Clean Air Initiative Asia (CAI-Asia) [59], which was generated from available datasets in 2005. All air quality stations were available for monitored important pollutants such as PM 10 and NO 2 in Thailand since the enhancement and conservation of the National Environmental Quality Act of 1992 [61]. PM stations for the GBD study were about 16 stations for representing entire areas in the Southeast Asia region [59], reflecting significant evidence of air pollutant concentrations and long distance correlation (i.e. regional scale).
PAF is an estimation of the proportion of cases in the entire study population that can be attributed to air pollution. It can illustrate the health impact gained if the exposure to the counterfactual level can be reduced. The PAFs in this study were between 7.6% and 16.9% (Table 3) depending on health outcomes. In another approach to assess exposure based on the same RR information [3], Fann et al. [62] estimated that the largest percentage (between 7% and 17% depending on the health outcomes) of mortality attributable to PM 2.5 was in southern California in the United States, using the Community Multiscale Air Quality (CMAQ) Modelling System [63] and health impact function. Anenberg et al. [20] also estimated the global burden of mortality due to PM 2.5 to be about 2% for cardiopulmonary and lung cancer mortality, and 7% for all-causes mortality using the global atmospheric chemical transport model [64]. In another related study based on the same CR [3,65], Ying Li et al. 2010 [66] estimated the disease burden attributed to particulate matter exposure in the United Arab Emirates (UAE) from surface air monitoring station data and the spatial interpolated modelling technique. Their estimates of attributable fractions for PM were represented spatially and ranged from 12% to 28% of the total all-causes mortality in the UAE in adults aged >30 years in 2007, or at approximately 545 excess deaths annually [66]. The all-causes mortality due to PAF of PM 10 and PM 2.5 in our study were approximately 3 and 8%and lower compared to Ying Li et al. The means of PM 10 concentration (μgm -3 ) for the UAE (90-665 μgm -3 ) were higher than those in Thailand (ranging between 20 and 84 μgm -3 ) because the UAE is situated in a desert region and severe dust storms occur in the Arabian Gulf region [66]. Moreover, the dispersion of pollutants from other continents may be another factor producing a high natural background of pollutants (e.g. PM 10 90 μgm -3 and PM 2.5 , 45 μgm -3 ).
According to a previous study on the mortality risk estimation due to air pollution in Thailand, the pollution mix, seasonality and demographics may be different from developed countries in Europe and North America [31]. We attempted to use available RR information from local epidemiological studies to reduce the bias caused by extrapolation of findings to another location [66]. Therefore, we used the RR information for PM 10 all-cause and respiratory outcome from a local study that investigated the association between effects of exposure to air pollution on mortality risks in Thailand [9]. For NO 2 , we estimated all-causes mortality and respiratory mortality based on evidences from a systematic review and meta-analysis on a global scale due to the lack of RR information at the local level. However, the burden of NO 2 showed the largest mortality contribution, and the high correlation between NO 2 and PM 2.5 (around 0.7-0.8) of meta-analysis still suggests the possibility that NO 2 effects could be due in part to confounding from particulate matter. Hence, future epidemiological studies about information on the RR for Thailand should be conducted to reduce bias and improve PAF estimation.
Our results may be underestimated, since GBD recommended O 3 as one of the indicators to quantify air pollution exposure associated with adverse health outcomes similar to those induced by PM (i.e. respiratory, cardiopulmonary diseases) [46,66]. Several studies [67,68] indicated that NO 2 contributed O 3 formation as a precursor with heavy traffic load, large population density and meteorological factors [69][70][71]. Moreover, mixtures of O 3 and NO 2 might react to form dinitrogen pentoxide (N 2 O 5 ), that could create a greater risk than either O 3 or NO 2 . Further studies should consider the analysis with the role of O 3 as a possible important effect on the health outcomes.
For the quality of information on the levels of mortality and causes of death, several studies stated that mortality statistics in Thailand were low quality, with 20-40% of deaths are registered with unknown or nonspecific causes in the past decade [56,72,73]. However, we used the best available mortality information from the study that had been initiated to verify cause of deaths (COD) reported by vital registration from the nation-wide VA study [53,56,74], and adjusted the completeness of the vital registration was based on the mid census SPC conducted by the National Statistical Office [57,58].
Our study might have some limitations and uncertainties. For the exposure assessment based on air quality monitoring station may depend on the location, density and distance of the monitoring network to nearby emission sources. In particular, the low number of measuring sites displayed in some regions (e.g. two and five stations in the north-eastern (about 160,000 km 2 ) and southern (about 70,713 km 2 ) regions, which may have some limitations in simulating the uniformly distributed annual ambient air pollution exposure on a large scale (e.g. national or regional scale. We recommend that empirical-based models at the national level are required to identify the priority sites of where new monitoring stations should be located to increase the air monitoring stations in a large population area [75], to improve the empirical-based estimation in future research.
Furthermore, monitoring stations with a measurement capacity for PM 2.5 remained limited at the national level at the time of this study [76]. Therefore, we recommend the use of a PM 2.5 /PM 10 ratio based on available local study and empirical information in WHO's EBD study [19] to estimate the exposure to PM 2.5 . As the remote sensing technique was used for estimating the surface PM 2.5 concentrations from satellite observations, remote sensingderived PM 2.5 measures have been found to be well correlated with actual ground-level PM 2.5 measurements [77,78]. Therefore, future research may consider remote sensing data combined with ground monitoring station data in Thailand for greater precision in assessing PM 2.5 exposure [47].
Another limitation was regarding the PAF estimation on a spatial scale based on two different grid resolutions of Pe and air pollutant concentration layer. Although, we used the resampling and interpolation technique in the raster algebra process [79], this may have produced some variation of grid and uncertainty from the estimation [66]. Further studies should employ a multi-spatial resolution approach [80,81] and/or consider the consistency of spatial resolution on air concentrations and the population of exposure distribution [48].
Finally, our findings indicate a significant health impact due to air pollution problems in Thailand and that a 20% reduction in air pollutants could reduce the number of annual deaths by about 160-7,425 per year. Therefore, the government should increase its effort and investment into controlling air pollution to achieve the SDGs. As previously stated, air pollution problems and their burden of disease are geographically specific. Thus, autonomy and the capacities of local authorities in managing their own problems are certainly required, as well as a national healthy public policy framework [82] to effectively deal with these problems.

Conclusions
This study aimed to quantify the magnitude and distribution of disease burden caused by ambient air pollution for policy-makers and planner by presenting an integrated exposure assessment, using a spatial interpolation model from empirical data, population distribution exposure and health impact function to estimate the national disease burden attributable to ambient air pollution. In addition, the GIS-based population exposure assessments for PAFs and the estimation of the number of deaths due to ambient air pollution exposure are useful for prioritising policy to reduce and prevent adverse health effects in Thailand. We hope that our findings offer a national estimate and benefit decision-making by stakeholders and policymakers to promote and develop air quality management and health co-benefit strategies to achieve the SDGs in the future.