figure a

Introduction

The escalating pandemic of type 2 diabetes presents an enormous public health challenge around the world. There were estimated to be 383 million diabetic individuals and around 1.4 million deaths due to diabetes worldwide in 2016 [1, 2]. These numbers are expected to continue to rise rapidly, especially in middle- and low-income countries [3]. Type 2 diabetes can lead to a series of chronic complications, including vision loss, renal diseases, stroke and cardiovascular diseases, that pose an overwhelming burden on healthcare systems [4,5,6]. Several traditional cardiovascular risk factors are also risk factors for the development of type 2 diabetes, including obesity, high blood pressure and unhealthy lifestyles and behaviours, which have all been well investigated [7].

Air pollution is the largest single environmental risk in the world and it has recently been regarded as a novel risk factor for cardiovascular diseases. The American Heart Association states that exposure to particulate matter (PM) with an aerodynamic diameter less than 2.5 μm (PM2.5) is causally associated with an increased risk of cardiovascular morbidity and mortality [8]. Given the close linkage between type 2 diabetes and cardiovascular disease, PM2.5 air pollution may also act as a risk factor for type 2 diabetes. Although a few studies have investigated long-term exposure to PM and the development, prevalence and mortality rate of diabetes [9,10,11,12], the results are inconsistent. Large-scale prospective cohort studies are necessary to provide stable results and precise estimates. Furthermore, most studies have been conducted in North America and Europe [13,14,15,16] and limited information is available from other regions, such as the WHO western Pacific and south-east Asia regions, where many countries are experiencing high levels of air pollution. At the same time, epidemics of type 2 diabetes are growing quickly in Asia. We therefore investigated the association of long-term exposure to PM2.5 with the development of type 2 diabetes in a large longitudinal cohort of 147,908 adults in Taiwan.

Methods

Study participants

This study was based on an ongoing large prospective cohort whose details were described previously [17,18,19,20]. In brief, this cohort study recruited more than 0.6 million participants between 1994 and 2014. A private firm, the MJ Health Management Institution, provided a standard medical screening programme. The participants were of Chinese descent residing in Taiwan. They were encouraged to visit the firm periodically through a paid membership and underwent a series of medical examinations at each visit, including anthropometric measurements, spirometry test, blood and urinary tests and imaging tests, and answered a standard self-administered questionnaire survey. This cohort is an open (dynamic) cohort with no end date. Each year there are around 20,000 new members recruited to the cohort, in addition to the revisits by existing members. Data generated from the medical examinations have been computerised since 1996. As of December 2014, the database contained around 0.59 million Taiwan participants and had counted 1.35 million medical visits. Around 43.5% of the participants had attended at least two medical visits (range 2–28 visits). Each participant was required to sign an informed consent form before participation. We obtained ethical approval for this study from the Joint Chinese University of Hong Kong, New Territories East Cluster Clinical Research Ethics Committee.

The participant selection in this study is presented in ESM Fig. 1. In total, 418,811 participants at least 18 years of age with fasting plasma glucose (FPG) measurements were recruited during 2001–2014, when data on PM2.5 concentration were available. We excluded 52,365 participants with incomplete information (2861 with PM2.5 data due to a missing address and 49,504 based on the covariates). We further excluded 15,008 participants with self-reported physician-diagnosed cancer or cardiovascular diseases at their first medical visit because of the possible effects of comorbidities on type 2 diabetes. Among the remaining 351,438 participants, 194,975 were excluded because they visited the institution and received medical examination only once. For those who had at least two medical examination visits, 5488 participants were excluded because they had diabetes (defined as FPG ≥7 mmol/l, or self-reported physician-diagnosed diabetes) at their first visit [21]. Because the development of type 2 diabetes is a chronic process, we further excluded 3067 participants with a follow-up duration of less than 2 years. Finally, 147,908 participants with 548,429 FPG measurements were included in the present analysis. Compared with those participants excluded because of attending only one medical visit, the 147,908 participants included in the present study had similar baseline distributions of general characteristics (ESM Table 1). The follow-up duration of the 147,908 participants ranged from 2 to 13.9 years (mean 6.7 years). The number of medical visits ranged from 2 to 23 with a median of 3.0. The mean visit interval was 1.2 years (SD 4.1).

Exposure assessment

We have described the detailed method for estimating PM2.5 exposure in previous publications [17, 19, 22, 23]. In brief, the ambient PM2.5 exposure at each participant’s address was estimated by a satellite-based spatial-temporal model with high resolution (1 × 1 km2) using the aerosol optical depth (AOD) data derived from the Moderate Resolution Imaging Spectroradiometer carried on US National Aeronautics and Space Administration satellites. The sample size for AOD data was around 300, which was similar to the general average level worldwide [24]. To address the issues of missing data and temporal limitation, we derived a correction factor using ground observations [23]. To calibrate the satellite-derived AOD data, we collected the ground observation of AOD from the aerosol robotic network (AERONET) in Taipei (EPA-NCU station, 24.97°N and 121.19°E), the capital city of Taiwan. Finally, we validated the model by comparing the estimated PM2.5 exposure with the monitoring data from more than 70 ground-level air pollution monitoring stations. The correlation coefficients for yearly average concentration ranged from 0.72 to 0.83 [17, 18].

The address of each participant (either residential or business) was noted during each medical visit so that the medical report could be mailed to them. Thus, any change of address was recorded. If a participant reported a change of his/her address in a follow-up medical visit, the PM2.5 concentration at the new address since the follow-up time point was applied in the data analysis. There were 29,032 (19.6%) participants who changed their address during this study. We geocoded each participant’s address into latitude and longitude data, which were used to calculate the address-specific yearly average PM2.5 concentration. The 2 year average concentration was then calculated based on the concentrations from the year of and the year before the medical examination as an indicator of long-term exposure to ambient PM2.5 air pollution.

Outcome measurement

Detailed information on the medical examination and quality control has been described in previous publications and in the Technical Reports published by the MJ Health Research Foundation [17, 18, 25]. An overnight fasting blood sample was taken in the morning and the plasma glucose level was measured enzymatically with a Hitachi 7150 analyser (Tokyo, Japan) if before 2005 or Toshiba C8000 analyser (Tokyo, Japan) if since 2005.

The health outcome in this study was incident type 2 diabetes. After the baseline assessment at the first visit, all the 147,908 non-diabetic participants were followed up, and incident type 2 diabetes was identified by medical assessment (defined as FPG ≥7 mmol/l, or self-reported physician-diagnosed diabetes) in subsequent visits [21]. The endpoint was the first occurrence of type 2 diabetes or the last visit if type 2 diabetes did not occur.

Contextual variables

We collected information on the participants’ demographic and socioeconomic characteristics, lifestyle and medical history by using a standard self-administered questionnaire at each visit. Height and weight were measured with participants wearing light indoor clothing without shoes. Seated blood pressure was measured using an auto-sphygmomanometer (CH-5000; Citizen, Tokyo, Japan). An overnight fasting blood sample was taken in the morning and a lipid profile was documented.

Based on previous literature [7, 26], we included the following factors as covariates in this study: age (years), sex (male or female), education (lower than high school [<10 years], high school [10–12 years], college or university [13–16 years] or postgraduate [>16 years]), smoking status (never, former or current), alcohol drinking (seldom [less than once per week], occasional [1–3 times/week] or regular [>3 times/week]), physical activity (defined as the product of metabolic equivalent value [MET; 1 MET = 1 kJ h−1 [kg bodyweight]−1] and duration of exercise [h] [27]: inactive [<3.75 MET-h], low [3.75–7.49 MET-h], medium [7.50–16.49 MET-h], high [16.50–25.49 MET-h] or very high [≥25.50 MET-h]), vegetable intake (seldom [<1 serving/day], moderate [1–2 servings/day] or frequent [>2 servings/day]), fruit intake (seldom [<1 serving/day], moderate [1–2 servings/day] or frequent [>2 servings/day]), occupational exposure to dust or organic solvents in the workplace (yes or no; as obtained by asking, ‘Are there any occupational hazards in your workplace?’), BMI (calculated as weight [kg] divided by the square of height [m]), hypertension (defined as systolic blood pressure ≥ 140 mmHg, diastolic blood pressure ≥90 mmHg or self-reported hypertension), dyslipidaemia (defined as total cholesterol ≥13.3 mmol/l, triacylglycerol ≥11.1 mmol/l or HDL-cholesterol <2.2 mmol/l), season of each visit (spring [March–May], summer [June–August], autumn [September–November], winter [December to February]) and calendar year at baseline.

Data analysis

We used Cox regression models with time-dependent covariates to analyse the associations between long-term exposure to ambient PM2.5 and the development of type 2 diabetes. The timescale used in the models was time-in-study (i.e. follow-up time). A crude model and three multivariable models were developed to compare the effects of covariates: Model 1 had no adjustment; Model 2 adjusted for demographic factors (age, sex and education), season and calendar year; Model 3 further adjusted for lifestyle factors (smoking status, alcohol drinking, physical activity, vegetable intake and fruit intake) and occupational exposure and Model 4 further adjusted for cardiovascular risk factors (BMI, hypertension and dyslipidaemia). All covariates were treated as time-dependent covariates except for sex. We estimated the HR with 95% CI as the risk of incident diabetes for PM2.5 quartiles (first to fourth quartile [μg/m3], respectively): <21.7, 21.7–<24.1, 24.1–<28.0, ≥28.0). We tested the linearity using likelihood ratio test and results showed that the associations deviated from linearity (χ2 = 35.2, p< 0.05). Therefore, we applied deciles to show the concentration–response associations between ambient PM2.5 and incident type 2 diabetes (first to tenth deciles [μg/m3], respectively: 5.7–<19.8, 19.8–<21.2, 21.2–<22.2, 22.2–<23.2, 23.2–<24.1, 24.1–<25.2, 25.2–<26.5, 26.5–<32.7, 32.7–<39.9 and 39.9–50.3).

We performed stratified analyses based on the following characteristics of the participants at baseline: sex (male vs female); education (<13 years vs ≥13 years); smoking status (never vs ever); alcohol drinking (seldom vs occasional/regular); physical activity (<7.5 MET-h vs ≥7.5 MET-h); BMI (<23 kg/m2 vs ≥23 kg/m2) [28]; hypertension (no vs yes) and dyslipidaemia (no vs yes). We stratified the participants into two subgroups according to the baseline cut off values of the aforementioned characteristics. We then performed data analyses separately within each subgroup using the Cox regression models with time-dependent covariates to derive the stratum-specific HR.

To test the stability of the associations, we further conducted sensitivity analyses by: (1) including participants with a follow-up of less than 2 years; (2) including only participants enrolled before 2005 whose FPG and lipids were measured with the Hitachi 7150; (3) including only the participants enrolled since 2005 whose FPG and lipids were measured with the Toshiba C8000; (4) excluding the participants with a business address to eliminate the potential misclassification of PM2.5 exposure due to different types of addresses; (5) using annual average PM2.5 concentration as an indicator for long-term exposure to PM2.5; (6) excluding those participants who were younger than 30 years old to better distinguish between type 1 and type 2 diabetes; (7) further adjusting for region of participant’s location (including five municipalities [Taipei, Taoyuan, Taichung, Tainan and Kaoshiung], ten counties [Hsinchu, Miaoli, Changhua, Nantou, Yunlin, Chiayi, Pingtung, Ilan, Hualien and Taitung] and one county-level city [Keelong]) to consider the effects of different regions; and (8) only including those participants with annual medical visit (interval of the medical visits ranged from 8 months to 16 months) to avoid delayed diagnosis of the disease.

All the statistical analyses were performed using R 3.3.2. (R Core Team, Vienna, Austria). The exposure and interaction effects were regarded as statistically significant at a two-tailed test level of 0.05 and 0.1, respectively.

Results

Table 1 shows the characteristics of all participants and participants with incident diabetes. The mean age of the participants was 38.3 years (SD 11.5 years) at baseline. The majority were well-educated, non-smokers and seldom drank alcohol; 4781 participants developed diabetes. At baseline, the participants who developed diabetes were generally older, had a lower level of education and were more likely to smoke and drink alcohol. They also had a higher prevalence of cardiovascular risk. The cumulative incidence was 3.2% with an incidence rate of 3.5 per 1000 person-years.

Table 1 Characteristics of the participants
Full size table

The locations of the participants are shown in Fig. 1. The participants mainly lived in the western part of Taiwan. In general, the south-western areas were the most heavily polluted and the middle and eastern areas were the least heavily polluted. The spatial pattern of exposure contrast throughout the island generally remained stable during the study period. The PM2.5 concentrations increased slightly from 2001 to 2004 (the mean 2 year PM2.5 was 24.8, 26.2, 28.7 and 29.6 μg/m3, respectively, for participants enrolled in 2001, 2002, 2003 and 2004) and then declined gradually from 2005 to 2014 (the mean 2 year PM2.5 was 27.4, 26.9, 26.9, 26.7, 26.8, 25.8, 25.6, 25.0, 23.7 and 24.4 μg/m3, respectively, for 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012, 2013 and 2014). The overall mean was 26.5 μg/m3 (SD 7.4 μg/m3) with an interquartile range of 21.7–28.0 μg/m3.

Fig. 1
figure 1

Spatial distribution of the participants in Taiwan, by year. The spatial distribution of the 147,908 participants at baseline by year and mean PM2.5 concentration at each location during that year. Circles indicate participant locations

Full size image

Table 2 shows the associations between ambient PM2.5 and the development of type 2 diabetes. Exposure to PM2.5 was significantly associated with a higher risk of incident type 2 diabetes. Compared with the participants exposed to the first quartile of ambient PM2.5, those exposed to the second, third and fourth quartiles of PM2.5 were associated with HRs (95% CI) of 1.28 (1.18, 1.39), 1.27 (1.17, 1.38) and 1.16 (1.07, 1.26), respectively, for incident type 2 diabetes after adjusting for a wide range of covariates. The concentration–response association is presented in Fig. 2. No obvious effect threshold was observed.

Table 2 Associations of long-term exposure to ambient PM2.5 with incident type 2 diabetes in Taiwanese adults
Full size table
Fig. 2
figure 2

The concentration–response association between ambient PM2.5 and incident type 2 diabetes in the Taiwanese participants. The data are presented as estimated HR (95% CI) associated with PM2.5 concentration deciles. The association was adjusted for age, sex, education, season, year, smoking status, alcohol drinking, physical activity, vegetable intake, fruit intake, occupational exposure, BMI, hypertension and dyslipidaemia. The PM2.5 range (μg/m3) of the first to tenth deciles, respectively, was: 5.7–<19.8, 19.8–<21.2, 21.2–<22.2, 22.2–<23.2, 23.2–<24.1, 24.1–<25.2, 25.2–<26.5, 26.5–<32.7, 32.7–<39.9 and 39.9–50.3

Full size image

The results of stratified analyses are presented in Table 3. Significant modified effects were observed for the categories of BMI and alcohol drinking. Long-term PM2.5 exposure had a stronger association with the development of diabetes in participants with BMI <23 kg/m2 or a habit of occasional or regular alcohol consumption. No significant modifying effects were observed for the other factors. Sensitivity analyses generally yielded similar results (ESM Table 2).

Table 3 Stratified analyses of the associations between PM2.5 and incident type 2 diabetes by covariates at baseline
Full size table

The associations between covariates and incident type 2 diabetes are presented in ESM Table 3.

Discussion

The results of this large prospective cohort study show that long-term exposure to ambient PM2.5 appears to be associated with a higher risk of developing type 2 diabetes after controlling for a wide range of covariates. The associations remain robust in the stratified and sensitivity analyses.

Our results are in line with those of some previous studies [16, 29,30,31]. Three studies, in Canada, Denmark and the USA, found that a 10 μg/m3 increase in PM2.5 was associated with a higher risk of incident diabetes with the HR ranging from 1.11 to 1.52 [16, 29, 30]. In a Hong Kong elderly population, a significant association with an HR of 1.15 per interquartile range (3.2 μg/m3) was also observed by Qiu et al [31]. The larger HR in the Hong Kong study was possibly due to its elderly participants who were potentially more vulnerable. There are limited cohort studies in Asia but a few large-scale cross-sectional studies showed that PM2.5 was significantly associated with higher risk of prevalence of diabetes [12, 32], supporting our findings. Several previous studies also found that traffic-related air pollution or other air pollutants, including NO2 and O3, were associated with incident diabetes [14, 33,34,35]. However, five previous studies did not find that PM significantly affected the development of diabetes [26, 35,36,37,38]. Many factors may contribute to this inconsistency, including the heterogeneity of the study populations, study regions, chemical components of PM, research methods and diagnostic criteria. Our study targeted a population in Taiwan, in which the PM2.5 levels were higher than those in North America and Europe. One advantage of this study is that the exposure contrast among our participants was relatively large. The accuracy of the exposure estimate may also play an important role in the inconsistency among the studies. Some previous studies estimated PM exposure based on the proximity of residences to fixed monitoring stations, with the same exposure level assigned to an entire community (district, county or city). Such community-level exposure assessment (ecological fallacy) may mask the true spatial variation or introduce misclassification of exposure, thus leading to the inconsistent results. Finally, the effect sizes of PM2.5 on incident type 2 diabetes are small. Thus, a large sample size is necessary to provide sufficient statistical power, yet many previous studies had relatively small sample sizes.

In this study, we categorised PM2.5 into quartiles. It seems the HR values decreased slightly in participants with higher quartile exposure (HR 1.28, 1.27 and 1.16 for the second, third and fourth quartile, respectively). We do not know the exact reasons for this phenomenon, but we speculate that the use of category variable might lead to a loss of information and an increase in uncertainty. Because the likelihood ratio test show that the association marginally deviated from linearity (χ2 = 35.2, p < 0.001), we used PM2.5 deciles to draw the concentration–response association (Fig. 2). The HR values decreased in the sixth to ninth deciles and jumped in the tenth decile. It is difficult to interpret this phenomenon but the non-linearity association might also be due to the heterogeneities of the populations in different cities/areas and other unidentified confounders (e.g. some factors might affect the health of people living in the same city but this may vary across the city). The high HR in the tenth decile of PM2.5 concentration could be partially explained by the larger PM2.5 concentration range (39.9–50.3 μg/m3) and the relatively smaller number of incident diabetes cases (there were 416 cases in the tenth decile, while the number in each of the first nine deciles ranged from 450 to 521). Nonetheless, further studies to illustrate the concentration–response relationship between PM2.5 and type 2 diabetes are warranted.

The biological mechanism underlying the association between long-term exposure to PM2.5 and the development of type 2 diabetes is not completely understood. Animal experiments have shown that PM2.5 can produce hypothalamic inflammation and induce metabolic disorders, including autonomic imbalance, visceral adipose inflammation, endothelial dysfunction, insulin resistance and overt diabetes [39, 40]. One in vivo study noted that long-term exposure to PM2.5 can further cause metabolic disorders by triggering the unfolded protein response and macrophage infiltration [41]. In addition, cardiovascular disease and diabetes may have similar mechanistic pathways (systematic inflammation and oxidative stress) [42]. Our previous study clearly showed that PM2.5 may induce systemic inflammation [17, 43]. Thus, the systematic inflammation and oxidative stress induced by PM2.5 may be the main mediators between PM2.5 exposure and diabetes by disrupting insulin signalling [44, 45].

We also explored the potential modifying effects of a range of factors. Although no significant modifying effects were observed for sex, education, smoking status, physical activity, hypertension and dyslipidaemia, statistical significance was observed for the modifying effect of alcohol drinking. Presently, however, there is little information on the modifying effects of alcohol drinking. BMI was another significant modifier in this study. Interestingly, the participants with a lower BMI had a higher risk of developing diabetes due to PM2.5 exposure, even though BMI is a significant risk factor for diabetes development. This phenomenon has also been observed in previous studies but the modifying effects were generally insignificant [29, 32, 33, 46]. Further studies are warranted to assess the different modifying effects.

This study has several important strengths. First, it targeted a large general population in Asia, where type 2 diabetes epidemics are growing quickly and people are generally experiencing serious air pollution. Second, it used a longitudinal study design, and most incidences of type 2 diabetes were identified by FPG measurements. In comparison with self-reported diabetes, the FPG test is a relatively time-efficient way to minimise diagnostic misclassification and lessen the likelihood of underestimating the incidence of type 2 diabetes [34]. The longitudinal study design also enabled us to account for the effects related to the change of PM2.5 exposure and a wide range of covariates. The associations did not change materially after including these covariates in the models. Third, the large sample size and the relatively long follow-up duration gave the study sufficient power to detect the small effects of ambient PM2.5 on the development of type 2 diabetes. The large sample size also allowed us to generate stable results and precise estimates. Finally, we used a novel model based on satellite-derived AOD data with high resolution (1 × km2) to estimate individual exposure at the participants’ addresses. This technology permitted us to overcome the spatial coverage limitation that typically occurs when using data obtained only from monitoring stations. The exposure data at individual level also enabled us to avoid ecological fallacies.

This study also has several limitations. First, we did not have information on indoor and gaseous pollutants. However, we included smoking as a covariate as it is an important source of household air pollution in developed economies. The generally high correlations between gaseous pollutants and PM2.5 suggest that we should analyse their effects separately [16]. Second, the PM2.5 exposure levels were calculated at the fixed addresses of the participants, and their daily activity patterns were not considered. More advanced technologies are needed for more accurate assessment of exposure in future studies. Third, we did not account for noise exposures due to the information being unavailable. Noise may be regarded as a potential risk of cardiovascular disease [47]. Fourth, it is difficult to distinguish between type 1 and type 2 diabetes in a large-scale epidemiological study with around 0.15 million participants. However, the targeted population were non-diabetic participants aged 18 years or above and therefore the majority of individuals who developed diabetes were likely to develop type 2 diabetes. The sensitivity analysis that excluded participants with a baseline age of <30 years yielded similar results (ESM Table 2), which further supported the association between PM2.5 and the development of type 2 diabetes. Fifth, the follow-up frequency and interval of medical examinations varied among the participants. Thus, it is difficult to identify the exact onset date of the disease for those participants with a long interval between medical examinations. However, the sensitivity analysis that only included the participants with annual medical examination yielded similar results. Finally, the participants were relatively healthy and were educated to a high level. Therefore, we should be cautious when generalising the results to other populations.

In conclusion, we found long-term exposure to ambient PM2.5 to be significantly associated with a higher risk of developing type 2 diabetes in a population from Asia, in which people in many of the region’s countries are generally experiencing high levels of air pollution and the prevalence of diabetes is rising rapidly. We advocate urgent strategies to reduce global air pollution that can aid in preventing the current pandemic of type 2 diabetes.