Differential Effects of Source-Specific Particulate Matter on Emergency Hospitalizations for Ischemic Heart Disease in Hong Kong

Background: Ischemic heart disease (IHD) is a major public health concern. Although many epidemiologic studies have reported evidence of adverse effects of particulate matter (PM) mass on IHD, significant knowledge gaps remain regarding the potential impacts of different PM sources. Much the same as PM size, PM sources may influence toxicological characteristics. Objectives: We identified contributing sources to PM10 mass and estimated the acute effects of PM10 sources on daily emergency IHD hospitalizations in Hong Kong. Methods: We analyzed the concentration data of 19 PM10 chemical components measured between 2001 and 2007 by positive matrix factorization to apportion PM10 mass, and used generalized additive models to estimate associations of interquartile range (IQR) increases in PM10 exposures with IHD hospitalization for different lag periods (up to 5 days), adjusted for potential confounders. Results: We identified 8 PM10 sources: vehicle exhaust, soil/road dust, regional combustion, residual oil, fresh sea salt, aged sea salt, secondary nitrate, and secondary sulfate. Vehicle exhaust, secondary nitrate, and secondary sulfate contributed more than half of the PM10 mass. Although associations with IQR increases in 2-day moving averages (lag01) were statistically significant for most sources based on single-source models, only PM10 from vehicle exhaust [1.87% (95% CI: 0.66, 3.10); IQR = 4.9 μg/m3], secondary nitrate [2.28% (95% CI: 1.15, 3.42); IQR = 8.6 μg/m3], and aged sea salt [1.19% (95% CI: 0.04, 2.36); IQR = 5.9 μg/m3] were significantly associated with IHD hospitalizations in the multisource model. Analysis using chemical components provided similar findings. Conclusion: Emergency IHD hospitalization was significantly linked with PM10 from vehicle exhaust, nitrate-rich secondary PM, and sea salt–related PM. Findings may help prioritize toxicological research and guide future monitoring and emission-control polices. Citation: Pun VC, Yu IT, Ho KF, Qiu H, Sun Z, Tian L. 2014. Differential effects of source-specific particulate matter on emergency hospitalizations for ischemic heart disease in Hong Kong. Environ Health Perspect 122:391–396; http://dx.doi.org/10.1289/ehp.1307213


Introduction
Over the past decades, epidemiologic evi dence has linked ambient particulate matter (PM) pollution to increased cardio vascular morbidity and mortality (Dominici et al. 2006;Peng et al. 2009). Of the cardio vascular end points, ischemic heart disease (IHD) is a major public health concern. IHD is defined as a narrowing of the coronary ves sels that supply blood to the heart muscle. It was the leading cause of death worldwide in 2008 and the second leading cause of death in 2011 in Hong Kong, which had a popula tion of around 7 million and a daily average of 12 IHD deaths that year (Department of Health HKSAR 2013). Evidence from the United States and Europe of increases in IHD events after acute exposure to elevated PM concentrations has been convincing (Dominici et al. 2006;Forastiere et al. 2005;. Dominici et al. (2006) estimated that an average 10μg/m 3 reduc tion in PM 2.5 (≤ 2.5 μm in aerodynamic diameter) in 204 U.S. counties would pre vent > 1,500 IHD hospitalizations per year. However, previous studies in Hong Kong did not observe associations between PM and IHD hospitalizations (Wong CM et al. 2002;Wong TW et al. 1999) or mortality upon adjusting for gaseous pollutants (Wong TW et al. 2002). Heterogeneity in findings may reflect the fact that PM is a complex mix ture of particles that vary in physical attri butes, chemical composition, solubility, and emission sources (Pope and Dockery 2006).
Growing research emphasis has been placed on PM sources and chemical composi tion (Health Effects Institute 2002;National Research Council 2004). Because PM sources generate mixtures of air pollutants with dif ferent physicochemical compositions, the source might affect the relative toxicity of PM. This hypothesis is supported by toxico logical evidence suggesting that PMinduced biologic effects can depend on the zone of origin (e.g., the industrial zone; Alfaro Moreno et al. 2002). Currently, the majority of studies have associated IHD, especially myocardial infarction, with trafficrelated pol lution exposures estimated using surrogate pollutants (e.g., PM 2.5 mass, carbon mon oxide, nitrogen dioxide) or directexposure data (e.g., time spent in traffic) (D'Ippoliti et al. 2003;Lanki et al. 2006b;Peters et al. 2004). Despite these findings, it has been a challenge to quantitatively assess the impacts of multiple PM emission sources on IHD. Associations with PM sources have been inconsistent across existing studies. Although some studies have reported associations of trafficrelated and/or combustiongenerated PM with increases in repolarization, inflam matory markers, and ST segment depressions among IHD patients (Lanki et al. 2006a;Yue et al. 2007), others have reported that IHD hospitalizations were not linked with trafficrelated particles or other PM sources (Halonen et al. 2009;Lall et al. 2011).
In Hong Kong, although research on PM pollution and health outcomes has been active since the late 1990s, specific PM chemical components and sources responsible for the adverse effects have rarely been investigated. In the present study, we took advantage of the PM 10 (≤ 10 μm in aerodynamic diameter) speciation data that have been available for over a decade to identify contributing sources to PM 10 mass using a source apportionment model and then used those data to estimate the acute effects of PM 10 sources on daily emergency IHD hospital admissions.

Data. The Hong Kong Environmental
Protection Department has been collect ing 24hr filter samples of PM 10 regularly at six general and one roadside air quality monitoring stations since 2001 (Yuan et al. 2013). These monitoring stations were inter spersed in different districts of Hong Kong. We included only data from the six general stations that are not in direct proximity to traffic, industrial sources, buildings, or resi dential sources of emissions from the burn ing of coal, waste, or oil. These stations serve volume 122 | number 4 | April 2014 • Environmental Health Perspectives to capture the air quality that the general population is exposed to on a regular basis. Twentysix PM 10 chemical components were speciated from the filter samples via various analytical methods as described in detail pre viously (Yuan et al. 2013). We included spe ciation data from between 1 January 2001 and 31 December 2007 in the present study. The PM 10 sampling frequency was on average every 6th day, with each station operated on a distinct sampling schedule. On a particular day, there might be no or multiple samples taken across the stations. Overall, 71% of the study days were covered by measurements from at least one station. We obtained daily mean temperature and relative humidity from the Hong Kong Observatory for the same study period.
We acquired daily counts of emergency hospital admissions for between 1 January 2001 and 31 December 2007 from the Hong Kong Hospital Authority (Wong TW et al. 1999). Data were coded according to the International Classification of Diseases, 9th Revison (ICD9;World Health Organization 1977). Hospitalizations for IHD (ICD9 codes 410-414) were extracted to construct the time series. Hospitalizations due to influenza (ICD9 code 487) were extracted and treated as a potential confounder in the regression analysis.
We removed the stationspecific influ ence on the resultant concentrations of each PM 10 source by a) computing the annual mean concentration (X i ) for each monitor ing station i, b) subtracting the annual mean from the daily mean concentration for station i on each sample day j (X ij ), c) adding the annual mean of all stations (X) to the result ing centered values (X ij -X i ) for each station and sampling day to produce X´´i j = X ij -X i + X, and d) taking the average of X´´i j over all stations (Wong CM et al. 2001). The final PM 10 sources time series contained nonmiss ing territorywide mean concentrations of PM 10 sources for 1,805 days (71% of the 2,556 total days), which is about 5 days/week. All pollutant concentrations were expressed inmicrograms per meter cubed, except for EC and OM, which were reported in microgram of carbon per meter cubed.
Generalized additive models with log link and Poisson error were used to estimate the associations between PM 10 sources and emer gency IHD hospital admissions (Hastie and Tibshirani 1990). We adopted a priori model specification to guide the selection of degrees of freedom (df) for timevarying variables: smoothing splines with 8 df per year for time trend, 6 df for current day temperature and previous 3days moving average, and 3 df for current day relative humidity and previous 3days moving average Peng et al. 2009). We included dummy variables for day of week, public holidays, and influ enza epidemics (Wong CM et al. 2002).
We investigated the possible lag distribu tion of associations with each PM 10 source for exposures on the same day (lag 0 ) and for daily exposures on the previous 1-5 days (lag 1 to lag 5 ). However, we focused primarily on the 2day moving average of exposure on the same day and the previous day (lag 01 ) a priori based on previous studies (Wong CM et al. 2002, 2008. Furthermore, we conducted multi source analyses to estimate mutually adjusted effects of PM 10 sources on emer gency IHD hospitaliza tions (Ostro et al. 2011). To minimize multi collinearity, we used backward elimination with an exclusion criterion of p > 0.10 to select PM sources Figure 1. PM 10 source profiles, indicated by explained variations that estimate how much a source explains the variation of a particular chemical component.

Vehicle exhaust
Soil and road dust  to include in the final multi source model while controlling for time trend, seasonality, meteorological conditions, calendar effects, and influenza epidemics. Pearson's correla tions were used to summarize the relation ships between sourceapportioned PM 10 . PM 10 "tracer" components, which are char acterized as the typical components that are exclusively or largely derived from a particular source, were also examined, and those trac ers that are found specifically in the sources included in the final multi source model were further tested in a separate multi pollutant model to validate the multi source findings. A smoothing function with 3 df was applied to graphically describe the relationships between sources and IHD hospitalizations while adjusting for time varying confounders. For sensitivity analyses, we repeated the timeseries analyses after either imputing source concentrations for the days without samples from any stations (751 days) by lin ear interpolation using the na.approx func tion in the R zoo package or by replacing the missing data with nonmissing measurement values from the previous day. Moreover, we evaluated the impact of alternative df values (5-12) for time trend on the risk estimates. All estimates were reported as the percent increase [(relative risk -1) × 100%] in daily emergency IHD hospital admissions for an inter quartile range (IQR) increment in pollu tant concentrations. Where appropriate, 95% confidence intervals (CIs) were calcu lated. We performed all timeseries analy ses in the statistical environment R Software, version 2.15.0 (R Foundation for Statistical Computing, Vienna, Austria).

Results
We identified 8 PM 10 sources, namely vehicle exhaust, soil/road dust (e.g., from exposed soil, unpaved roads), regional combustion, residual oil combustion (e.g., fuel emissions from marine vessels), fresh sea salt, aged sea salt, secondary nitrate, and secondary sulfate. Figure 1 shows the estimated PM 10 source profiles, depicted as explained varia tions that indicate the relative contribution of each source to the variation of a given chemical component (Paatero and Tapper 1994). For instance, vehicle exhaust emission accounted for 80% of the variation in EC. Regional combustion emission was identi fied as a composite of two sources that could not be further separated. They were wood/ biomass burning [based upon the abundance (i.e., a large explained variation) of K + ] and coal combustion in power plants and indus trial facilities in the adjacent Pearl River Delta region (based upon the abundances of As, Cd, Pb, and Zn in the source profile, which can not be further separated) (Yuan et al. 2013). Table 1 summarizes the levels of PM 10 pol lution, weather conditions, and IHD hos pital admission counts. Between 2001 and 2007, the daily average concentration for PM 10 in Hong Kong was 55.8 ± 32.5 μg/m 3 . Secondary sulfate accounted for the largest fraction of total PM 10 mass (23.6%), fol lowed by vehicle exhaust (15.1%) and sec ondary nitrate (14.9%). The mean daily average temperature and relative humid ity were 23.6°C and 78.3%, respectively (Table 1). During the study period, there were 76,659 hospitalizations for IHD (30 ± 7 admissions per day).
Singlesource models of singleday exposure lags showed similar patterns of associations for most of the PM 10 sources, in that IHD hospitalizations were positively associated with exposure on the same day (lag 0 ), maximal for lag 0 or lag 1 , and low est at later lags (lag 4 -lag 5 ) (Figure 2). At lag 01 (Figure 3A), the source that was most strongly associated with IHD hospitaliza tions was secondary nitrate [2.89% increase (95% CI: 1.83, 3.95); IQR = 8.6 μg/m 3 ], followed by vehicle exhaust [2.35% (95% CI: 1.24, 3.47); IQR = 4.9 μg/m 3 ] and regional combustion [2.26% (95% CI: 0.98, 3.55); IQR = 11.7 μg/m 3 ], after adjusting for time trend, seasonality, meteorological conditions, calendar effect, and influenza epidemics. Significant positive associations were also found for particles originated from  Correlations between sourceapportioned PM 10 were nil to moderate. The highest cor relation coefficient, 0.67, was between regional combustion and secondary sulfate, followed by 0.59, between regional combustion and secondary nitrate (Table 2). Backward elimina tion resulted in a multi source regression model that included vehicle exhaust, aged sea salt, and secondary nitrate sources only ( Figure 3B). All other sources, although statistically signifi cant in singlesource models, were eliminated from the final multi source model on the basis of p > 0. 10 Figure 3C).
We examined the concentration-response relations for vehicle exhaust, aged sea salt, and secondary nitrate in a multi source model. We observed moderate positive relationships over the IQRs of source concentrations, except for aged sea salt, where a neutral relationship was seen (data not shown). The risk estimates were not sensitive to alternative timeseries models in which we imputed missing data (data not shown). Varying df for time trend (5-12 per year, data not shown) did not sub stantially change the regression results either.

Discussion
Research directly delineating the health impacts of PM emission sources is relatively limited. Most studies rely on ambient con centrations of a PM chemical component as a surrogate of the combined exposure to one source (Stanek et al. 2011). This a priori selection can be complicated when interpret ing the results because many components are emitted from numerous sources and the same component may not serve as tracer to the same source at different locations (Sarnat et al. 2008). We joined a small but growing number of epidemiologic studies to conduct source apportionment analysis and quan titatively estimate the associations between multiple PM sources and health outcome. Early shortterm air pollution studies con ducted in western countries identified some associations between PM sources and mor tality (Cakmak et al. 2009b;Halonen et al. 2009;Ito et al. 2006;Laden et al. 2000;Mar et al. 2000Mar et al. , 2006Ostro et al. 2011), and gradually, researchers have also linked certain PM sources to hospital admissions (Andersen et al. 2007;Cakmak et al. 2009a;Halonen et al. 2009;Lall et al. 2011;Sarnat et al. 2008). Overall, these studies have reported some evidence suggesting that PM sources representing traffic/motor exhausts, regional/ secondary sulfate, and coal/oil combustion may be more toxic than other PM sources, as summarized in a recent review by Stanek et al. (2011). Nonetheless, there is insufficient evidence to draw more specific conclusions across studies. Because emission sources of air pollutants vary not only temporally but also geographically, studies on source apportioned PM mass under different atmospheres are needed to improve our understanding of PMrelated health effects.
To our knowledge, this is the first Asian study to investigate the health impacts of multiple PM sources. We estimated the associations of shortterm exposure to source apportioned PM 10 mass with emergency IHD hospital admissions in Hong Kong, a coastal urban city on the boundary region of Asian continent and Pacific Ocean. In con trast to previous studies conducted in New York City, New York (USA), and Helsinki, Finland, that reported no associations of PM 2.5 sources with IHD hospital admissions (Halonen et al. 2009;Lall et al. 2011), we observed significant associations between IQR increases in several PM 10 sources and daily IHD hospitalizations for singleday lag peri ods up to 5 days prior. Differences between our findings and those of previous studies might be related to the longer study period (7 years), larger combined sample size in the present study, as well as the differences in pollution compositions and population sus ceptibility between cities. Although precise patho physiological mechanisms connecting ambient air pollution with IHD remain to be determined, it is commonly hypothesized that PM sources may trigger and/or enhance the formation of reactive oxygen species that induce inflammation, the formation of athero sclerotic plaques, and vaso constriction, resulting in reduced oxygen supply of heart tissues, and thereby leading to IHD (Lawal and Araujo 2012; Peters 2011).   Table 1 for individual IQR values for sources, and the IQR for EC (tracer for vehicle exhaust), Na + (for aged sea salt), and NO 3 -(for secondary nitrate) was 1.6, 1.2, and 3.4 μg/m 3 , respectively. Secondary nitrate in PM 10 (per 8.6 μg/m 3 ) was associated with the largest increases in IHD hospitalizations at lag 01 in both single source and multi source models. However, the estimated association of secondary sulfate diminished after adjusting for other sources, and secondary sulfate was dropped from the final multi source model based on the back ward elimination criterion (data not shown). This was somewhat surprising considering that secondary sulfate accounted for the larg est fraction of total PM 10 in Hong Kong. Whereas most studies that examined these associations found that sulfaterich second ary PM was more strongly associated with mortality and hospital admissions than nitrate (e.g., Ito et al. 2006;Mar et al. 2000Mar et al. , 2006Sarnat et al. 2008), a few studies have reported that nitrate in PM 2.5 , rather than sulfate, was significant predictor of mortality (Fairley 1999;Ostro et al. 2011). Secondary nitrate and secondary sulfate, respectively, derive largely from the oxidation of nitro gen oxides and sulfur dioxide emitted from combustion of fossil fuels. Although both are acidic in nature, their strength of acidity var ies greatly depending upon the cityspecific interactions between local emissions, regional transports, and meteorological conditions (Schlesinger and Cassee 2003). Although ani mal toxicological evidence is inconclusive, Kelly and Fussell (2012) hypothesized that acidic aerosols may lower the pH within the airways by depositing hydrogen ions, thereby triggering adverse reactions. In China, strong economic growth and high total energy con sumption have led to substantial increases in anthropogenic nitrogen and sulfur emis sions over the past decades Lu et al. 2010). Studies showed that emis sions of nitrogen oxides and sulfur dioxide in the adjacent Pearl River Delta region due to rapid industrialization and urbanization have been the dominant contributors to secondary nitrate and secondary sulfate in Hong Kong through regional transportation (Guo et al. 2009;Yuan et al. 2013). Our finding on sec ondary nitrate is of particular importance for lending urgency to policy makers, particularly in developing economies, regarding both local and regional emission control and reduction of gaseous pollutants.
We found that an IQR increment (4.9 μg/m 3 ) in the 2day moving average con centration of PM 10 from vehicle exhaust was associated with a 1.91% estimated increase (95% CI: 0.70, 3.13%) in IHD hospital admissions after adjusting for other statisti cally significant sources. In accordance with these results, EC as a chemical tracer of vehicle exhaust (largely from diesel engines) was also significantly associated with IHD hospitalization risk. Vehicle exhaustrelated PM refers to combustion derived particles that primarily accumulate in the fine frac tion of PM 10 . Previous epidemiologic studies have reported that mobile sources PM 2.5 and EC are stronger predictors of overall cardio vascular outcomes than other sources and components (Cakmak et al. 2009a(Cakmak et al. , 2009bLall et al. 2011;Mar et al. 2006;Ostro et al. 2011;Sarnat et al. 2008), which is consis tent with our finding of an association of IHD with vehicle exhaust PM. Plausible biological mechanisms include elevated lev els of inflammatory biomarkers, impaired endotheliumdependent vasodilation, and promotion of STsegment depression (Dales et al. 2007;Lanki et al. 2006a;Yue et al. 2007). In Hong Kong, where road density was among the highest in the world at 254 vehicles per kilometer of road in 2009, expo sure to trafficrelated air pollution is ubiqui tous (World Bank 2012). These findings on vehicle exhaust particles stress the importance of the continuous reduction of overall traffic and related emissions and the reconfiguration of urban environments to reduce personal exposure to traffic.
We observed that aged sea salt was associ ated with an increased risk of IHD hospital izations after adjusting for other sources. Sea salts are most abundantly found in the coarser fraction of PM 10 . Whereas Mar et al. (2006) reported that sea salt was consistently associ ated with elevated cardio vascular and total mortality in Atlanta, Georgia (USA), across the various sourceapportionment analyses in an intermethod comparison study, most epidemiologic studies that estimated this association found no relationship between sea salt and health outcomes (Andersen et al. 2007;Gent et al. 2009;Lanki et al. 2006a;Ostro et al. 2011). Lee et al (2008) and Zhuang et al. (1999) suggested possible rela tions between sea salt and secondary nitrate because nitric acid may react with marine particles to form coarse mode nitrate along coastal areas. However, we observed nullto weak correlations between secondary nitrate and sea salts. This association should be investigated further.
Our findings add to the existing literature in several ways. First, we examined air pollu tion association with a specific cardio vascular end point, as opposed to a broad composite end point of different cardio vascular events, to provide better insight into the plausible biologic mechanisms. Second, with nearly 80,000 IHD hospital admissions over 7 years, our study was well powered to detect statisti cally significant associations. Moreover, this was one of the few epidemiologic studies that focused on exposure to sourceapportioned PM 10 , whereas most available studies were based on sourceapportioned PM 2.5 . This allowed us to identify adverse associations of not only sources that primarily generate finer mode PM 10 , but also those that produce coarser mode PM 10 .
Although we provided evidence of the health impacts of several PM 10 sources in Hong Kong, these findings should be inter preted with caution. Whereas the backward elimination procedure was used to identify a subset of predictors with the most statistically significant relationship with IHD hospitaliza tion, this approach might not guaran tee a truly "best" reduced model (Breiman 1996). The importance of PM 10 sources (e.g., regional combustion, secondary sulfate) excluded from the final multi source model should not be diminished because the statisti cal elimination procedure does not indicate or account for biological importance. Another limitation of this study was the every6th day sampling scheme for the PM 10 speciation data, resulting in nearly onethird of study days without samples from any stations. Exposure misclassification error might exist; however, our risk estimates were insensitive to alternative interpolation methods (data not shown). Moreover, PM from local emissions (e.g., vehicle exhaust, soil/road dusts) tend to have more error than PM from regional sources (e.g., secondary PM), given their higher spatial heterogeneity (Ito et al. 2004). Such issues of representativeness associated with PM sources may hinder the interpreta tions of the relative strengths of the observed associations in monitor based studies of ambient PM pollution.

Conclusion
We report evidence that PM 10 from vehicle exhaust, nitraterich secondary PM, and sea salt-related PM were significantly associ ated with elevated IHD hospitalization risks in Hong Kong. This study joins a growing body of literature to report evidence of adverse effects of sourceapportioned PM mass, which would help prioritize research on the biologic mechanisms linking PM pollution to car diac events and guide future monitoring and emission control polices.