Establishment of a City-Based Index to Communicate Air Pollution-Related Health Risks to the Public in Bangkok, Thailand

Abstract

An Air Quality Health Index (AQHI), a health risk-based air pollution index, was constructed to communicate to the public their health risks due to exposure to air pollution in Bangkok, Thailand. This AQHI was built by analyzing the association between total excess respiratory disease-related deaths and individual air pollutants, using a time-series analysis of daily data from 2010 to 2019. We used Poisson regression in a generalized additive model, with natural cubic smooth splines to analyze the data and controls for other common variables (time, temperature, relative humidity, day of the week, and public holidays). The regression coefficients of these variables were then employed to establish the suitability of this AQHI for Bangkok. The results indicated that a 10-unit increase in particulate matter (PM₁₀), fine particulate matter (PM_2.5), ozone (O₃), or nitrogen dioxide (NO₂) was statistically associated with increased respiratory disease mortality. The coefficients of these four pollutants were then adopted in the construction of an AQHI for Bangkok. Compared with the currently used Air Quality Index (AQI), the AQHI was a more effective indicator in communicating multiple air pollution-related health risks to the public in Bangkok.

1. Introduction

Exposure to air pollution has been clearly linked to an increase in negative health effects, such as respiratory problems, with higher rates of morbidity and mortality among those exposed to such pollution than those with less exposure to air pollution [1]. According to the World Health Organization (WHO), the greatest burden of outdoor air pollution, with 91% of the 4.2 million premature deaths, is in low- and middle-income countries in the western Pacific and southeast Asia regions [2]. In the case of Bangkok, Thailand, its air pollution situation is frequently severe during the dry season, when the atmospheric conditions are stable and the winds are light [3,4,5]. Health impacts due to exposure to air pollution in the city remain a key public health concern [6]. Due to the impact of air pollution on health, Thailand provides simple reports about air quality conditions to the public to warn people about the current air pollution levels in specific areas. There are also recommendations and guidelines for individuals to reduce their exposure to air pollution. The tool used for these reports is called the Air Quality Index (AQI), which has been used in numerous countries around the world. The AQI is calculated based on the concentration of common air pollutants (particles with an aerodynamic diameter of ≤2.5 μm (fine particulate matter, PM_2.5); particles with an aerodynamic diameter of ≤10 μm (particulate matter, PM₁₀); carbon monoxide (CO); nitrogen dioxide (NO₂); sulfur dioxide (SO₂); and ozone (O₃). The index with the highest value is then used to represent the air quality at that time [7,8]. However, the AQI only takes into account the single highest index, despite the fact that the health impact of air pollution depends on the multitude of air pollutants to which people are exposed. This tool is therefore of limited use for capturing the actual risk to the public due to exposure to air pollution [9].

To address this weakness, researchers undertook an initiative to construct a health-based index, which they referred to as the Air Quality Health Index (AQHI) [10]. The AQHI informs the public about the air pollution situation in connection with the associated health risks. Similarly to the AQI, the AQHI categorizes the health risks due to exposure to air pollution into various levels (low, medium, high, and very high) and provides recommendations based on these levels and for different population groups (e.g., the general public and sensitive individuals such as children, the elderly, and the unwell). The AQHI is applied differently in individual countries, given that the index was established based on associations between local air pollution and health effects related to exposure to air pollution, using time-series analyses. This requires extensive information about air pollutant concentrations, meteorological conditions, and health factors (morbidity and mortality) in the locations of interest. Worldwide, the AQHI has so far been officially implemented in Canada, Hong Kong, and Mexico City [9,11,12]. AQHIs have been developed in many countries around the world, including Canada [10], Sweden [13], China [14,15,16,17], and Mexico [9]. However, to the best of our knowledge, no case studies of using an AQHI to represent air pollution in southeast Asia have yet been conducted.

To investigate the effects of air pollution on health, previous studies took all non-accidental health conditions into account [10,15]. Respiratory diseases have been found to be strongly associated with exposure to air pollution [18,19,20]. The effects of air pollution can impact both mortality and morbidity in different contexts [21,22]. Several studies have suggested that respiratory disease mortality is a promising endpoint for analyses in cities facing severe episodes of air pollution [15,23]. In the case of Bangkok, a link between daily air pollution levels and mortality due to respiratory diseases has also been observed [24,25]. In China, researchers constructed an AQHI based on short-term mortality risk for the cases of Shanghai and Guangzhou, where serious problems with air quality have been observed for many years [15,17]. However, these AQHIs relied on all non-accidental mortality, not just respiratory disease-linked mortality.

In this study, we examined associations between concentrations of multiple air pollutants and respiratory mortality outcomes based on daily data collected over a 10-year period, using a generalized additive model (GAM) for time-series analyses. We then used the findings of these analyses to construct an AQHI suitable for Bangkok. The validity of the AQHI was examined, and scaling of the AQHI for public communication was introduced.

2. Materials and Methods

2.1. Overall Process

Based on the Canadian approach [10], we analyzed the impact of exposure to air pollution on mortality rates by means of a GAM in a time-series analysis [26]. The variables employed in this technique included forecasting variables (air pollutants), confounding variables (relative humidity and temperature), and indicator variables (calendar day, weekday, weekend, and public holidays). The number of respiratory disease-related deaths in Bangkok was used as a dependent variable. The pollutants that were statistically associated with mortality, independent of other air pollutant criteria, were then selected to calculate the AQHI. The AQHI was then validated by comparing the observed data with those from the currently used AQI [15].

2.2. Air Pollution and Meteorological Data

Bangkok, the capital of Thailand, is located in the central area of the tropical country. In 2021, Bangkok had a population of about 6 million, and it has an area of 1568.7 km². In Bangkok, air quality monitoring stations are located at roadside and general ambient areas. Roadside monitoring aims to assess the impacts of vehicle traffic emission on air quality, while ambient sites are representative for the public exposure to outdoor air pollution. Hence, in order to construct an AQHI, we collected air pollution concentration and meteorological data from seven permanent, ambient-air monitoring stations (operated by the Pollution Control Department, PCD), each located in one of seven districts of Bangkok, namely Bang Kapi, Bang Na, Din Daeng, Phaya Thai, Thon Buri, Wang Thonglang, and Yan Nawa. We obtained daily mean data for the following pollutants: PM₁₀, SO₂, NO₂, O₃, and CO, from 1 January 2010 to 31 December 2019. Monitoring of PM_2.5 data began in 2015 at the five stations located in Bang Kapi, Bang Na, Din Daeng, Phaya Thai, and Wang Thonglang; thus, we collected PM_2.5 data from these stations for the 2015–2019 period. The meteorological data comprised the daily mean temperature and daily mean relative humidity. Once all the data had been rechecked for completeness, we averaged them to represent a single daily condition for Bangkok.

2.3. Mortality Data

Thailand’s Ministry of Public Health reports the number of deaths per day in Bangkok, classified according to the International Statistical Classification of Diseases and Related Health Problems, 10th edition (ICD-10). Previous studies have established AQHIs using the ICD-10 categories A00–R99, which include all deaths due to disease but not those caused by accidents. Given that air pollution has a direct impact on the respiratory system [19,20], and that previous studies [25,27] have indicated a notable association between air pollution and mortality due to respiratory diseases in Bangkok’s population, we selected only the ICD-10 categories J00–J99 to represent the deaths caused by respiratory diseases. We collected these data for the same time period as the air pollution and meteorological data.

2.4. Statistical Analysis

We used a GAM in a time-series study to investigate associations between time-varying pollution exposure with time-varying mortality. This was a type of ecologic study because it analyzed daily population-average mortality and exposure levels. If the air pollution effects are small and the mortality outcomes are rare, the bias from ignoring the data-aggregation across individuals should be small.

The GAM was used to estimate effects associated with exposure to air pollution while accounting for smooth fluctuations in mortality that could confound estimates of the pollution effect. The GAM framework was chosen because it is based on a model where relationships between predictors and the dependent variable follow smooth patterns that may be linear or nonlinear.

In the context of a regression model for time-series analysis, GAM is a generalized linear model with a linear predictor, involving the sum of the smooth functions of covariates. In general, the model has a structure as follows:

$$g(\mathrm{E}\left(Y_t\right))={\beta }_0+{\beta }_1{x}_{1,t}+{\beta }_2{x}_{2,t}+\dots +{\beta }_p{x}_{p,t}+f_1\left(x_{p+1,t}\right)+\dots +f_m\left(x_{p+m,t}\right)$$

(1)

where $Y_t$ is a response variable at time $t$, $\mathrm{E}\left(Y_t\right)$ is the expected value of $Y_t$, and $g(\mathrm{E}(Y_t))$ is the link function that links the expected value to the predictor variables, $x_1,x_2,\dots ,x_p$. The parameters ${\beta }_1,\dots ,{\beta }_p$ are regression coefficients. The distribution of $Y_t$ is an exponential family distribution with mean ${\mu }_t$ and scale parameter $\varphi$, and $f_1,\dots ,f_m$ are smooth functions of the covariates $x_{p+1},\dots ,x_{p+m}$, respectively [28].

When the response variable is a count of some occurrences, such as the number of deaths in a day, we assume that the response $Y_t$ has a Poisson distribution, with the expected count of $Y_i$ being $\mathrm{E}\left(Y_t\right)={\mu }_t$. That is $Y_t ~ poisson\left({\mu }_t\right)$, for $t=1,2,\dots ,n$. The link function is usually the (natural) log; that is, $g({\mu }_t)=\log {\mu }_t$.

Smooth functions, $f_j$, of covariates are often constructed using regression splines, loess smoothers, or smoothing splines with a smoothing parameter $\lambda$. Regression splines can be expressed as a linear combination of a finite set of basis functions that do not depend on the dependent variable $Y_t$, which is practical for prediction and estimation. They fit a piecewise polynomial to the range of covariate $z_j$, partitioned by knots (k knots produce k + 1 piecewise polynomials). The polynomials can be of any degree d, but are usually in the range [0, 3] and most commonly degree 3. A cubic spline with k knots can be modeled as follows:

$$Y_t=b_0+b_1{x}_{jt}+b_2{x}_{jt}^2+b_3{x}_{jt}^3+b_{4}h\left(x_{jt},{\xi }_1\right)+b_{5}h\left(x_{jt},{\xi }_2\right)+\dots +b_{k+3}h\left(x_{jt},{\xi }_K\right)$$

(2)

where $h\left(x_{jt},\xi \right)$ is a truncated power basis function per knot, defined as

$$h\left(x_{jt},\xi \right)=\{\begin{array}{ll}{\left(x_{jt}-\xi \right)}^2& \mathrm{if} x_{jt}>\xi \\ 0& \mathrm{otherwise}\end{array}.$$

To fit a cubic spline to a data set with k knots, we performed least squares regression with an intercept and k + 3 predictors. This amounts to estimating a total of k + 4 regression coefficients; for this reason, fitting a cubic spline with K knots uses k + 4 degrees of freedom (df).

A natural spline is a regression spline with additional boundary constraints: the natural function is required to be linear at the boundary (in the region where x is a spline smaller than the smallest knot or a spline larger than the largest knot). This additional constraint means that natural splines generally produce more stable estimates at the boundaries. A natural cubic spline has $k+4$ to 5 df due to the constraints at the end points. The way to decide how many df to use is to try out different numbers of knots and see which produces the best-looking curve. A somewhat more objective approach is to use cross-validation [29].

In this study, the Poisson regression for time-series analyses was applied to predict the daily number of deaths caused by respiratory disease based on air pollutant concentration in Bangkok. Two indicator variables, weekdays (1 = weekdays, 0 = weekends) and public holidays (1 = public holidays, 0 = others), were taken into account in the model. Natural cubic splines were employed to control the influence of season, temperature, and relative humidity with the df [15,30,31,32]. The GAM for respiratory disease-related deaths takes the following additive form:

$$\begin{array}{ll}\log \left(\mathrm{E}\left(Y_t\right)\right)=& {\beta }_0+{\beta }_i{x}_{i,t-l}+b_{1}weekda{y}_t+b_{2}holida{y}_t\\ & +ns\left(calendartime,d{f}_1\right)+ns\left(temperature,d{f}_2\right)\\ & +ns\left(humidity,d{f}_3\right)\end{array}$$

(3)

where $\mathrm{E}\left(Y_t\right)$ is the average number of respiratory disease-related deaths at day $t$, $x_{i,t-l}$ is the mean concentration of the ith air pollutant (PM_2.5, PM₁₀, SO₂, NO₂, O₃, or CO) at day $t-l$, and $l$ is the Lag of the pollution exposure, which is generally restricted to 0 to 7 days. The parameter ${\beta }_i$ represents a coefficient estimate of the ith air pollutant, describing the change in the logarithm of the population average mortality count per unit change in the $i\mathrm{th}$ air pollutant at Lag $t$.

The natural cubic smoothing splines $ns\left(x,df\right)$ denote a smooth function of a covariate x (calendar time, temperature, or humidity) with the degree of freedom df. For studies of non-accidental mortality, a reasonable choice of df 1 is to set 7 df per year, so that little information from time scales of longer than two months is included when estimating ${\beta }_i$ [15,30,31,32]. This choice largely eliminates expected confounders resulting from seasonal changes in mortality and other longer-term trends. Similarly, we applied 6 and 3 df for the entire period in natural cubic smoothing splines to adjust the current day’s temperature and relative humidity, respectively [15].

We estimated the coefficients ${\beta }_i$ of the ith air pollution from Equation (3) and then obtained six individual models for six pollutants and selected those air pollutants that were significantly correlated with mortality to calculate the AQHI. The statistical analysis was performed using R software (version 4.1.0) with the mgcv package.

2.5. Constructing the AQHI Equations

To determine the mortality risk from the concentration of air pollutants, we considered the relative risk (RR), which is the ratio of the average mortality due to exposure to air pollutants relative to non-exposure. From Equation (3), the average mortality $\mathrm{E}\left(Y_t\right)$ can be calculated by taking the exponential function on each side. Then, the RR can be calculated as:

$${\mathrm{RR}}_{i,t-l}=\exp \left({\beta }_0+{\beta }_i{x}_{i,t-l}\right)/\exp \left({\beta }_0\right)=\exp \left({\beta }_i{x}_{i,t-l}\right).$$

(4)

The excess risk (ER) of mortality is given by the RR minus 1. Thus, the percentage ER of mortality was calculated by:

$${\mathrm{ER}}_{i,t-l}=100\ast \left(\exp \left({\beta }_i{x}_{i,t-l}\right)-1\right)$$

(5)

where ${\mathrm{ER}}_{i,t-l}$ is the percentage daily increase in respiratory disease-related deaths caused by exposure to ith air pollutants at Lag l.

For the purpose of numerical simplification, the percentage daily increase in respiratory disease-related deaths (${\mathrm{ER}}_{i,t-l}$) at day $t$ was scaled up by multiplying each value by 10 and dividing by the maximum value obtained from the summation of ${\mathrm{ER}}_{i,t-l}$ of selected air pollutants in the period of time from day 1 to day q, as shown in (7). This resulted in an unbounded AQHI with 10 levels, 1–10. To simplify reporting, the AQHI can be rounded to an integer.

$$\mathrm{AQHI}=10/c\ast {{\displaystyle \sum }}_{i=1\dots p}100\ast \left(\exp \left({\beta }_i{x}_{i,t-l}\right)-1\right),$$

(6)

$$\mathrm{where} c=ma{x}_{t=1\dots q} ({{\displaystyle \sum }}_{i=1\dots p}100\ast \left(\exp \left({\beta }_i{x}_{i,t-l}\right)-1\right)$$

(7)

2.6. Evaluation of the AQHI Validity

To evaluate the validity of the AQHI, the associations between the AQHI and the AQI with daily respiratory mortality rates were examined. We calculated the regression of the daily index values of both indices against daily mortality rates due to respiratory diseases, using the time-series analysis for the same period as included when constructing the AQHI. We then compared the associations of both indices to analyze the validity of the AQHI by presenting their regression coefficient estimates. We also examined variations in the daily AQHI and air pollutants included in the AQHI over a 6-month period (1 July to 31 December 2019) to demonstrate the applicability of the AQHI.

3. Results

3.1. Descriptive Analysis

The descriptive statistics for daily respiratory disease-related deaths, air quality, and meteorological conditions in Bangkok are summarized in

Table 1. Descriptive statistics of the daily number of respiratory disease-related deaths, air pollution, and meteorological data in Bangkok.

Variable	Mean	SD	Min	Max	P25	P50	P75	WHO Guideline
Number of deaths based on ICD-10 (J00–J99) b	10.4	3.5	0.0	27.0	8.0	10.0	13.0
Air pollution
PM2.5 (μg/m3) a	24.8	13.6	5.4	90.0	14.9	20.5	30.9	15
PM10 (μg/m3) b	39.7	18.7	12.1	156.1	26.5	34.7	47.5	45
SO2 (μg/m3) b	6.2	3.1	0.4	30.7	4.0	5.5	7.8	40
NO2 (μg/m3) b	38.6	15.9	7.4	118.6	27.0	34.3	47.7	25
O3 (μg/m3) b	59.5	25.0	5.8	170.6	41.0	56.2	75.3	NA
CO (ppm) b	1.1	1.7	0.2	23.2	0.5	0.7	0.9	3.5
Meteorology
Temperature (°C) b	29.1	1.8	18.2	34.3	28.1	29.3	30.3
Relative humidity (%) b	70.2	8.8	37.8	96.6	64.5	69.9	75.9

^a Daily data from 1 January 2015–30 June 2019. ^b Daily data from 1 January 2010–30 June 2019. P₂₅, P₅₀, and P₇₅ refer to the 25th, 50th, and 75th percentiles, respectively. NA indicates that the daily number is not available in the WHO guidelines. SD, standard deviation.

. The number of deaths was on average 10.4 per day. For PM_2.5, there were only 25% of the total number of days (the 25th percentile) below the WHO guideline [33]. In other words, the majority (75%) of its daily concentrations were higher than the guideline. Note that maximum CO concentration was an outlier which appeared in its long-term data; its effect on construction of AQHI was negligible. The mean daily temperature and relative humidity in Bangkok (29.1 °C and 70.2%, respectively) were close to the mean climate statistics for Thailand between 1981 and 2010 (mean temperature and relative humidity of 28 °C and 70%, respectively) [34].

3.2. Formation of the AQHI

The relationships among air pollution species were analyzed by calculating Spearman correlation coefficients; this nonparametric statistical test was used because these variables were not normally distributed [35]. The results in

Table 2. Spearman correlation coefficients for the air pollutants.

Spearman Correlation	PM2.5 (μg/m3)	PM10 (μg/m3)	SO2 (μg/m3)	NO2 (μg/m3)	O3 (μg/m3)	CO (ppm)
PM2.5 (μg/m3) a	1.000	0.913 **	0.469 **	0.649 **	0.644 **	0.478 **
PM10 (μg/m3) b		1.000	0.332 **	0.650 **	0.649 **	0.467 **
SO2 (μg/m3) b			1.000	0.475 **	0.139 **	0.236 **
NO2 (μg/m3) b				1.000	0.407 **	0.631 **
O3 (μg/m3) b					1.000	0.240 **
CO (ppm) b						1.000

^a Daily data from 1 January 2015–30 June 2019. ^b Daily data from 1 January 2010–30 June 2019. ** The correlation coefficient is significant at the 0.01 level.

show that all pairs of air pollutants were statistically correlated at significance levels of 0.01. The strongest positive correlation was between PM_2.5 and PM₁₀ (r = 0.913), followed by PM_2.5 and NO₂ (r = 0.649), and PM_2.5 and O₃ (r = 0.644).

Table 3. Poisson regression coefficient estimates (and their confidence intervals, CI) of air pollution (per 10-μg/m³ increase).

Lag Days	Poisson Regression Coefficient $\widehat{\mathit{\beta }}$ (95% CI)
	PM2.5 a	PM10 b	O3 b	SO2 b	NO2 b	CO b
Lag0	0.0187 ** (0.0047, 0.0327)	0.0188 *** (0.0118, 0.0258)	0.0169 *** (0.0116, 0.0223)	0.0298 (−0.0137, 0.0732)	0.0193 *** (0.0109, 0.0276)	−0.0192 (−0.0970, 0.0586)
Lag1	0.0261 *** (0.0113, 0.0409)	0.0177 *** (0.0104, 0.0252)	0.0110 *** (0.0058, 0.0162)	0.0271 (−0.0171, 0.0713)	0.0209 *** (0.0122, 0.0297)	−0.0362 (−0.1144, 0.0420)
Lag2	0.0165 * (0.0016, 0.0314)	0.0163 *** (0.0088, 0.0237)	0.0114 *** (0.0064, 0.0165)	0.0243 (−0.0195, 0.0681)	0.0189 *** (0.0105, 0.0273)	−0.0294 (−0.1074, 0.0487)
Lag3	0.0213 ** (0.0070,0.0357)	0.0129 *** (0.0056, 0.0201)	0.0088 *** (0.0039, 0.0137)	0.0422 (−0.0012, 0.0857)	0.0161 *** (0.0079, 0.0243)	−0.0597 (−0.1384, 0.0189)
Lag01	0.0185 * (0.0041, 0.0329)	0.0197 *** (0.0125, 0.0269)	0.0204 *** (0.0147, 0.0261)	0.0397 (−0.0071, 0.0864)	0.0230 *** (0.0144, 0.0316)	−0.0320 (−0.1134, 0.0494)

^a Daily data from 1 January 2015–30 June 2019. ^b Daily data from 1 January 2010–30 June 2019. * Significance level at 0.05, ** significance level at 0.01, and *** significance level at 0.001. The numbers in bold font indicate the highest regression coefficient for the air pollutants.

shows the results of a Poisson regression analysis from the GAM shown in Equation (3) to predict the logarithm of the average number of respiratory disease-related deaths at day $t$ at the ith air pollutant concentration (per 10-μg/m³), $x_{i,t-l}$, at Lag l = 0, 1, 2, 3 (denoted as Lag1, Lag2 and Lag3, respectively) and moving averages of the ith air pollutant concentration (Lag01). It should be noted that these estimates describe the change in the logarithm of the average number of respiratory disease-related deaths per 10-μg/m³ change in air pollution on the same day, so the increase in the average number of respiratory disease-related deaths can be calculated by the exponential function of ${\widehat{\beta }}_i$. Similarly, increases in the concentration of these air pollutants also significantly affected the increase in the average number of respiratory disease-related deaths at all Lags.

We selected air pollutants where their coefficient estimates (${\widehat{\beta }}_i$) showed the strongest mortality effects to calculate the ER and AQHI at day $t$. The strongest effects indicate the highest respiratory mortality risk due to exposure to air pollution. Therefore, the coefficients of air pollutants fall in Lag01 were chosen to construct the AQHIs [15,25,30]. From the results shown in Table 3, the strongest effects of PM₁₀, O₃, and NO₂ on respiratory mortality were observed at Lag01. Therefore, ERs due to a 10-μg/m³ increase in PM₁₀, O₃, and NO₂ concentrations were responsible for 1.97% (95% CI, 1.25−2.69%), 2.04% (95% CI, 1.47−2.61%), and 2.30% (95%CI, 1.44−3.16%) increases in respiratory mortality, respectively. Similarly, the ER for PM_2.5 was 2.61% (95% CI, 1.13−4.09%) as its highest coefficient was found at Lag1. From Table 3, the strongest significant coefficients of air pollutants, except PM_2.5, were found at Lag01. Since a strong variation of air pollutants data was found on a daily basis, the moving average of air pollutants (Lag01) could help to reduce the variation.

Since PM_2.5 and PM₁₀ may originate from the same sources [36,37,38], and they were strongly correlated (r = 0.913, Table 2), we suggested taking into account these two pollutants separately, with other significant air pollutants, in constructing the AQHI. This approach was in line with other studies [10,15]. Using the regression coefficient estimates for 1-μg/m³ PM_2.5, PM₁₀, NO₂, and O₃, derived from Table 3, the AQHIs for PM₁₀ (AQHI PM₁₀) and PM_2.5 (AQHI PM_2.5) can be obtained from the following equations:

AQHI PM₁₀ = 10/100 *(100 * (e^0.00197*PM₁₀ − 1 + e^0.00204*O₃ − 1+ e^0.00230*NO₂ − 1))

(8)

AQHI PM_2.5 = 10/73 * (100 * (e^0.00261*PM_2.5 − 1 + e^{0.00204 *O}₃ − 1 + e^0.00230*NO₂ − 1))

(9)

where 100 represents the daily maximum value of the sum of mortality ERs associated with PM₁₀, O₃, and NO₂ throughout the entire period, and 73 represents the daily maximum value of the sum of mortality ERs associated with PM_2.5, O₃, and NO₂ throughout the entire period. However, the equations should be applied based on the availability of local air-quality data. We propose the use of AQHI PM₁₀ (Equation (8)) if local air-quality monitoring does not provide PM_2.5 data. AQHI PM_2.5 (Equation (9)) should be employed if both PM_2.5 and PM₁₀ data are available, given that PM_2.5 poses a greater health threat than PM₁₀ [39,40]. Since CO and SO₂ were not statistically significantly associated with respiratory mortality (Table 3), they were not considered in the AQHI equations.

3.3. Validity of the AQHI

The results shown in

Table 4. Poisson regression coefficient estimates (and their 95% confidence intervals, CI) of respiratory disease-related deaths associated with the AQHI and the AQI (per 10-μg/m³ increase) ^a.

Respiratory Disease-Related Mortality	AQHI PM10	AQHI PM2.5	AQI
Lag0	0.2603 (0.1814, 0.3393)	0.2922 (0.1951, 0.3893)	0.0086 (0.0059, 0.0113)
Lag1	0.3260 (0.2126, 0.4393)	0.2425 (0.1472, 0.3378)	0.0058 (0.0032, 0.0085)
Lag2	0.3056 (0.1939, 0.4174)	0.2162 (0.1223, 0.3101)	0.0062 (0.0036, 0.0088)
Lag3	0.2401 (0.1325, 0.3476)	0.1776 (0.0873, 0.2680)	0.0047 (0.0022, 0.0072)
Lag01	0.3926 (0.2802, 0.5050)	0.4427 (0.3066, 0.5791)	0.0091 (0.0063, 0.0119)

^a Daily data from 1 January 2010–30 June 2019. All had significance levels at 0.001. The numbers in bold font indicate the highest regression coefficient for the indices.

indicate that both daily AQHI PM₁₀ and AQHI PM_2.5 had significant associations with daily respiratory disease-related mortality over the nine-year period (1 January 2010 to 30 June 2019). The strongest estimates were found at Lag01, where the coefficient estimates for AQHI PM₁₀ and AQHI PM_2.5 were 0.3926 (95% CI, 0.2802 to 0.5050) and 0.4427 (95% CI, 0.3066 to 0.5791), respectively. For the association of AQI, the estimates were also significant and found to be strongest at Lag01; however, they were much weaker compared with those of the AQHI.

We also examined the performance of AQHIs when applied in practice by plotting their levels against air pollutant concentrations during wet and dry episodes in Bangkok during 2019 (Figure 1). This was based on the approach of Canadian AQHI development [10]. AQHI PM₁₀ and AQHI PM_2.5 were both relatively well aligned with daily variations in air pollution. However, compared with AQHI PM_2.5, AQHI PM₁₀ showed fewer changes in relation to fluctuations in pollution, but was still able to reflect the differences in concentration levels. Overall, the AQHIs enabled the capture of important changes in daily air pollution over long periods, particularly during the dry season when serious air quality issues are frequently present.

3.4. Banding the AQHI for Air Quality Reports Aimed at the Public

There have been several methods proposed for categorizing an AQHI to conveniently communicate health risks to the public. For instance, Canada [10] categorizes the AQHI of 10 Canadian cities based on the frequency distribution of daily maximum AQHIs, whereas Hong Kong [14] and Guangzhou [17] use the summation of daily excess risks (%ER) to determine AQHI risk levels by considering short-term exposure to air pollution, as recommended by the WHO as a cut-off to represent very high risk. Given that the AQHIs of Hong Kong and Guangzhou are based on single cities (as is the case in our study), we adopted this method for determining the risk level for Bangkok. Our AQHI was banded into four ranks, as shown in Figure 2, consisting of “low health risk” (bands 1–3), “moderate health risk” (bands 4–6), “high health risk” (bands 7–9), and “serious health risk” (band 10 and above).

We developed our methodology following those established for AQHIs in Hong Kong [14] and Guangzhou [17]. For AQHI PM₁₀, the current WHO guidelines [33] indicate a cut-off of 37.5 µg/m³ for PM₁₀ (24-h mean, interim third target third), 25 µg/m³ for NO₂ (24-h mean), and 100 µg/m³ for O₃ (8-h mean). Therefore, using Equation (5), the %ER is 44.47%, giving a “high health risk”. We then designated the %ER value that was 1.5-times higher than the “high health risk” %ER (i.e., 66.70%) as a “serious health risk”. The cut-off AQHI value between “low health risk” and “moderate health risk” was the %ER that was 50% less than the “high health risk” (22.24%). Similarly, in the case of AQHI PM_2.5, as the WHO [33] established a cut-off of 37.5 µg/m³ for PM_2.5 (24-h mean, interim third target third), with NO₂ and O₃ being the same values as for AQHI PM₁₀, using Equation (5), the %ER is 38.83%. This was designated as a “high health risk”. A %ER of 1.5-times higher than this value indicates a “serious health risk” (58.25%), while 50% (19.42%) is the cut-off AQHI value, classified as a “low health risk” and a “moderate health risk”.

Based on the %ERs above, we used a statistical method known as the class interval method to band the risks into 10 levels, as shown in Figure 2. It can be seen that around 60% of days were in the “moderate health risk” band for both AQHI PM₁₀ and AQHI PM_2.5. Around 14% of days fell within the “high health risk” and “serious health risk” bands for AQHI PM₁₀. This was similar to AQHI PM_2.5, for which approximately 18% of days were ranked into these two risk categories.

It should be noted that we selected the values specified in the interim third target of the WHO guidelines as a cut-off for PM₁₀ and PM_2.5 to generate the AQHI banding. This was based on a new 24-h mean PM_2.5 standard recently designated by the PCD using the values in this interim target. For NO₂ and O₃, we proposed the use of the final target in the WHO guidelines as they have not yet been specified as 24-h means in the present national standard or in the same interim target for particulate matter pollutants. In addition, the methodologies used in Hong Kong [14] and Guangzhou [17] provided a specific AQHI band for vulnerable groups (children and the elderly) using an adjusted %ER derived from the ratio of the median %ER for these groups. This was not within the scope of our study.

4. Discussion

4.1. Associations between Air Pollution and Respiratory Disease-Related Deaths

Globally, observational epidemiological studies have indicated that estimates of the impact of air pollution on mortality outcomes are more consistent and comparable than those for morbidity outcomes. This could be because morbidity rates vary according to the indicators used (hospital admissions, outpatient and emergency room visits, etc.) and socioeconomic contexts such as the health system and income levels [15,23,41,42]. In Bangkok, a link between daily air pollution concentrations and respiratory disease mortality has been observed [24,25,27]. We found that PM_2.5, PM₁₀, O₃, and NO₂ concentrations at Lag0, 1, 2, 3, and Lag01 were significantly correlated with respiratory disease-related deaths in Bangkok (Table 3). These pollutants were therefore included when constructing the AQHI for Bangkok. The present study reasonably reflects the current situation in relation to Bangkok’s air quality, where such pollutants are of concern to the Thai government [43]. However, our findings diverged from those of a Chinese study involving an AQHI in Shanghai [15], which indicated that PM_2.5, PM₁₀, and NO₂ were linked with mortality. One of the main reasons for this difference in findings was that, in the Shanghai study, all non-accidental causes of death (ICD-10 categories A00-R99) were included in the analysis, while our study considered only deaths due to respiratory diseases. Moreover, there may have been differences in environmental conditions, public health, socioeconomic status, and other factors.

4.2. Excess Risk of Mortality

The highest excess risk of mortality from exposure to PM_2.5 in Bangkok was 2.61%, as calculated using Equation (5). This shows that the percentage increase in respiratory disease-related deaths was higher than that found in a study involving 272 cities in China, which reported an ER of 0.29% (95% CI, 0.17–0.42%) [30], and in a study conducted in New England (USA), with an ER of 0.41% (95% CI, 0.16–0.65%) [44]. However, it was lower than that reported in a study of cities across Brazil, which found the association of PM_2.5 with respiratory disease-related deaths during wildfire episodes was 5.09% (95% CI, 4.73–5.44%) [45].

For PM₁₀, the ER in the current study was 1.97%, which is lower than that reported in a study conducted in Ahvaz, Iran [46], in which the ER was 2.7% (95% CI, 0.20–5.10%) [24]. Interestingly, while ambient PM₁₀ air pollution in Bangkok has decreased over the years, the ER value of PM₁₀ in our study (from the 2010 to 2019 data) was still higher than that reported by Vichit-Vadakan et al. [25], who used data from 1999 to 2003 and determined an ER in Bangkok of 1.0%. The reasons may be related to an increase in the population of Bangkok [3,47] over the passing years, together with changes in individuals’ behaviors, such as spending time outdoors (e.g., motorcycling, outdoor recreation), particularly on days with high levels of pollution.

For gaseous air pollution, our results demonstrated an association between O₃ and respiratory disease-related deaths that was higher than that reported in Jiangxi, China, with an ER of 1.04% (95% CI, 0.04–1.68%) [48], and in Oporto, Portugal, with an ER of 1.05% (95% CI, 0.45–2.57%) [49]. For NO₂, a recent study of Chinese cities reported an ER of 1.2% (95% CI, 0.9–1.5%) [50], lower than that found in our study. Conversely, a study conducted in 10 Italian cities found an ER of 3.48% (95% CI, 0.75–6.29) [51], higher than our result. The differences in the associations may be due to a variety factors, including different air pollution and meteorological conditions in different regions, as well as differences in people’s level of education, socioeconomic status, and so on [52,53,54].

In Bangkok, traffic and open burning of biomass are the main sources of particulate matter [6,43]. O₃ is a secondary gas formed in the air through interactions between nitrogen oxides (NOx) and volatile organic compounds (VOCs) under the influence of solar radiation [55]. Previous research conducted in Bangkok has indicated that O₃ has health impacts and still occurs at a high level [56,57]. Many countries have been looking for solutions due to the significant associations of these pollutants with a variety of respiratory diseases that have serious health outcomes [58]. Internationally, the WHO has recently updated its guidelines for ambient air pollution because the adverse health effects caused by such pollutants have seen a marked increase.

4.3. Establisment of the AQHI

In this study, we developed an AQHI for Bangkok based on a comprehensive investigation of long-term data on the associations between common air pollutants and respiratory disease-related mortality. Based on the results, we have presented an AQHI for Bangkok that includes PM_2.5 or PM₁₀, O₃, and NO₂. There is currently no consensus about which pollutants should be taken into account when constructing AQHIs for various health outcomes and geographical areas. The AQHIs in Canada and Mexico City were based on three pollutants, including PM_2.5, NO₂, and O₃ [9,10,59]. The AQHI for Shanghai included PM₁₀ or PM_2.5 and NO₂ [15]. AQHIs for Hong Kong, Guangzhou, and Ningbo all included PM_2.5, SO₂, NO₂, and O₃ [14,17,60]. These variations were mainly due to the differences in local air pollution and the health information used in constructing local AQHIs [17].

Considering the collinearity and health effects of PM_2.5 and PM₁₀, and consistent with the studies in Canada [10] and in Shanghai, China [15], we therefore proposed two AQHI models: AQHI PM₁₀ and AQHI PM_2.5. In practice, this might be an advantage as it provides options where either PM_2.5 or PM₁₀ information is incomplete due to the malfunctioning of measurement systems, or when the construction of standardized monitoring stations is in progress.

Previously, researchers have established AQHIs based at city and national levels. At the city level, Cromar et al. [9] and Chen et al. [15] constructed AQHIs for Mexico City and Shanghai, respectively. Stieb et al. [10] and Du et al. [16] developed AQHIs at the national level for Canada and China, respectively. It is reasonable to assume that the establishment of local AQHIs could provide more effective information to the public than national level AQHIs. In this study, we established an AQHI for Bangkok. This work originated from the findings of a comprehensive epidemiological study which showed remarkably different associations between health impacts and air pollution in Bangkok compared with Thailand’s other main cities [27].

We also investigated the AQHI’s validity by comparing the association magnitudes of the AQHI and the current AQI when estimating daily respiratory disease-related mortality risks (Table 4). The results revealed that the AQHI performed much better at estimating mortality rates than of the AQI did. Our results agreed with previous studies [15,16,60], where AQHIs were found to be better indicators for predicting mortality risks than AQIs. This may be explained by the fact that AQIs only consider the single pollutant with the greatest AQI value, based on piecewise linear functions [17,61], whereas AQHIs consider the influence of multiple pollutants on population health risks [17]. Therefore, AQHIs represent a more suitable indicator to inform the public about health effects due to air pollution.

Suggestions for avoiding the negative health effects associated with air pollution have also been provided to the public in accordance with the risk levels. An AQHI can serve as an effective air-quality reporting system for air quality management in areas of interest. Moreover, it can be used as an indicator to help a city assess its achievements toward United Nations Sustainable Development Goals (SDGs) associated with the health impacts of air pollution (e.g., SDG target 3.9.1, which proposes a substantial reduction in mortality and morbidity due to air pollution) [22,62].

Our study had some strengths. First, it represents the first AQHI established for a southeast Asian city, where air pollution-related health burdens continue to exist among citizens due to high levels of urbanization. This is one of the WHO’s main areas of concern [2]. Second, we used long-term data to select the air pollutants that are potentially linked to respiratory disease-related deaths to be included in the calculation of the AQHI [60]. We also introduced the option of using AQHI PM₁₀ when PM_2.5 data are not available due to a lack of suitable monitoring stations. This is a common situation in cities in low- and middle-income countries. Lastly, we also evaluated the validity of the AQHI compared with the currently used AQI and established a banding system to conveniently communicate the risks to the public. This provided comprehensive information to support the establishment of an AQHI as the new air quality indicator for the city.

4.4. Limitations

Given that an AQHI relies on air pollution and health data, it must be revised regularly based on air pollution and health situations that change over time. In addition, an AQHI is applicable only in areas where local data have been used to create the equations. Therefore, any researchers who intend to develop an AQHI for application in their own location should consider the completeness of their air quality and health information. It is unclear whether an AQHI can be generalized to different areas of a country with varying levels and patterns of air pollution [17]. We suggest AQHIs are only developed for cities where there is concrete scientific evidence for a high linkage of air pollution and health effects. It is possible to create a general health-based index using coefficients obtained from various locations [10,16], but this approach was not examined in our study. However, it is not advised to use such an index in place of current systems that call for necessary actions based on categories of outdoor pollution levels, such as closing schools when air quality exceeds ambient air quality standards [9]. In addition, it is not only common air pollutants that affect population health but also other hazardous air pollutants such as VOCs and polycyclic aromatic hydrocarbons (PAHs), which have been linked to potential health issues [58,63]. Due to limitations in measuring frameworks, it was not possible to include these pollutants to calculate the AQHI. Based on these various factors, the AQHI we developed can be used to present overall health effects to the public to help prevent them from being exposed to polluted air.

5. Conclusions

This study found strong associations between daily air pollutant (PM_2.5, PM₁₀, O₃, and NO₂) concentrations and daily respiratory disease-related deaths in Bangkok. We then established an AQHI based on these short-term associations. Our study demonstrates that the AQHI can be used as a tool to communicate air pollution-related health risks to the public, and we anticipate that this AQHI will be used officially to promote a better quality of life for citizens in Bangkok.