Risk/benefit tradeoff of habitual physical activity and air pollution on chronic pulmonary obstructive disease: findings from a large prospective cohort study

Our study applied data from the UK Biobank Cohort study [Application Number: 69550], an ongoing longitudinal cohort of over 0.5 million participants aged 40–69 years at baseline (2006–2010). The study was mainly conducted in urban areas of England, Scotland, and Wales through the UK National Health Services register [6]. All participants completed a series of baseline assessment, including socioeconomic factors, lifestyle and behavioral factors, and the history of medication and operations. Biological samples, such as blood and urine samples, were taken as well. Details of the UK Biobank protocols can be found elsewhere [14]. The UK Biobank study was approved by the North West Multi-centre Research Ethics Committee (06/MRE08/65). Informed consent was obtained from all participants [15].

Inclusion and exclusion procedures of the study population in the present study are outlined in Fig. 1. Of the 502,490 participants in the UK Biobank, 15,966 with prevalent COPD at baseline were excluded. We further excluded 149,230 participants due to missing data for at least one covariate in fully adjusted models, leaving 337,294 participants with complete covariate data. After further excluding 71,014 participants with missing PM_2.5 or PA, 266,280 participants were left in our final sample. As for participants with objectively measured PA, according to data exclusion criteria of accelerometers in previous studies [16, 17] and following the same selection procedures as participants with self-reported PA, a total of 59,948 participants were left in the subsample analysis.

Information on habitual PA was obtained by a self-administered questionnaire and wrist-worn accelerometer. Briefly, self-reported PA was ascertained with a modified version of the International Physical Activity Questionnaire (IPAQ), which includes six questions about duration and frequency of walking, moderate-intensity and vigorous-intensity exercise undergone in the last 4 weeks [18]. Each intensity was assigned a corresponding metabolic equivalent (MET, 1 MET = 1 kilocalorie per hour per kilogram of bodyweight): 3.3 for walking, 4.0 for moderate PA, and 8.0 for vigorous PA [19]. We then quantified PA of each participant by calculating minutes of MET each week (MET-min/week) based on the reported intensity, duration, and frequency of PA in 1 week. Afterwards, participants were classified into three groups based on a standard scoring criteria of International Physical Activity Questionnaire (IPAQ) [20]: low (< 600 MET-min/week), moderate (600 to 3000 MET-min/week), and high (≥ 3000 MET-min/week). The threshold of 600 MET-min/week is equal to reaching the recommended WHO guidelines for moderate-intensity PA (150 min per week) [21].

Furthermore, we adopted a subsample of about 100,000 participants with objectively measured PA using triaxial accelerometers (Axivity AX3, Newcastle upon Tyne, UK). Participants were chosen voluntarily and accelerometers (the Axivity AX3 wrist-worn triaxial accelerometer) were sent to their personal addresses [22]. They were required to wear accelerometers on their dominant wrists for seven consecutive days. Information on the accelerometer protocol, including data extraction and processing, has been documented in detail elsewhere [23]. Individuals were grouped into three groups based on the PA tertiles [low < 24.32 milli-gravity (mg), moderate 24.32 to 30.70 mg, and high ≥ 30.70 mg] [24].

The annual average concentration of PM_2.5 in 2010 was calculated using a land-use regression model (LUR) developed by the European Study of Cohorts for Air Pollution Effects (ESCAPE) project [25]. Details on the development and validation of the ESCAPE LUR models have been described elsewhere [26]. In brief, based on a range of predictive variables (such as traffic intensity, population, topography, and land use) derived from geographic information system (GIS), LUR models were used to calculate the spatial variation in air pollutant concentrations at individuals’ residential addresses provided at baseline. Leave-one-out cross-validation showed good model performance for PM_2.5 (cross-validation R²=77%) in the southeast England area (London/Oxford) [27]. The obtained PM_2.5 concentration in 2010 was used as a surrogate measure of long-term PM_2.5 exposure, given the fluctuation of temporal trend of PM_2.5 concentrations remained generally parallel during study period [28, 29].

Participants were followed up from enrollment till the first occurrence of COPD, death, loss of follow-up, or 31 December 2019, whichever came first. COPD cases were identified using self-reported information, primary care data, and hospital admission data through linkage to the UK National Health Services register [18]. COPD diagnoses were coded as J40-J44 according to the International Classification of Diseases version-10 (ICD-10).

Potential confounders were selected based on literature review a priori [6, 11, 30], including demographic characteristics (sex, age, ethnicity, etc.), lifestyle factors (smoking status, drinking status, and fruit and vegetable intake), and so on. Details of definitions of these covariates could be seen in the Additional file 1: Supplementary methods of covariates and 4-model analytical protocol [14, 21, 31‐33].

Cox proportional hazard regression models with follow-up time as time scale were applied to examine the associations of long-term PM_2.5 exposure and habitual PA with COPD incidence. STROBE cohort reporting guidelines were adopted as well [34]. We developed a 4-model analytical protocol by adding covariates (e.g., demographic characteristics, socioeconomic factors, lifestyle factors) incrementally. We investigated the independent associations of PA and PM_2.5 first by building Models 1, 2, and 3. Then Model 4 was constructed for mutual adjustment. Details of the 4-model analytical protocol can be found in Additional file 1: Supplementary methods of covariates and 4-model analytical protocol.

When exposures were included as continuous variables, hazard ratio, and 95% confidence interval (HR, 95% CI) were calculated for each inter-quantile range (IQR) increment in PM_2.5 (1.27 μg/m³) and per 600 MET-min/week in PA. For categorical exposures, effect estimates were calculated with reference to low level of PA or the first tertile of PM_2.5. The proportional hazard assumption was examined by plotting Schoenfeld residuals, and no evidence of serious violation was found. Concentration-response relationship between long-term PM_2.5 exposure and COPD incidence was examined using a spline term in Cox model, where PM_2.5 was treated as a continuous variable with a degree of freedom (df) of 4.

Subgroup analyses stratified by PM_2.5 tertiles (< 9.48 μg/m³, 9.48 to 10.27 μg/m³, and ≥ 10.27 μg/m³) or PA levels (low, moderate, and high) were also conducted using Model 3. Cochran-Armitage trend test was used to confirm the constant trend toward higher incidence of COPD with an increasing PA levels and PM_2.5 exposure levels [35].

We further investigated the potential interactions between PA and PM_2.5 exposure on both additive and multiplicative scales. For additive interaction, we firstly categorized study participants into three groups based on their PM_2.5 tertiles (low < 9.48 μg/m³, moderate 9.48 to 10.27 μg/m³, and high ≥ 10.27 μg/m³). Combined with PA levels (low, moderate, and high), we then created a new term with nine categories representing nine combinations (3 × 3) of PM_2.5 exposure levels and PA levels.

To test the additive interaction, we calculated the relative excess risk due to interaction (RERI) and their corresponding confidence intervals (CI) with the reference group of high level of PM_2.5 exposure and low level of PA. RERI measures the combined excess risk in both exposed group that is due to the interaction [36]. A RERI of 0 denotes no additive interaction (i.e., the combined excess risk is the sum of their individual excess risks, and the excess risk is calculated as HR-1), a RERI of more than 0 indicates positive interaction (i.e., the combined excess risk is more than the sum of their individual excess risks), and a RERI of less than 0 implies that the combined excess risk is less than the sum of their individual excess risks [37]. For example, a negative RERI value in the present study would indicate that the benefit of PA was attenuated by residential PM_2.5 exposure.

On multiplicative scale, we added a product term between PM_2.5 exposure and PA levels. Likelihood tests were applied to test the significance of interaction term by comparing the model with and without the interaction term. A p value of the interaction term less than 0.05 indicates a multiplicative interaction [38].

For objectively measured PA data extracted from Axivity AX3, we used average acceleration to represent the overall PA intensity and categorized participants into three levels [< 24.32 milli-gravity (mg), 24.32 to 30.70 mg, and ≥ 30.70 mg]. Detailed information of accelerometer-measured PA level can be found elsewhere [24, 39]. The analyzing protocol for objectively measured PA was the same with that of the self-reported PA.

We performed a number of sensitivity analyses to examine the robustness of our results. First, we additionally adjusted for baseline lung function, inverse distance to main road, and both [40, 41]. These two variables were weakly correlated with PM_2.5 and PA (Additional file 1: Table S1). Second, instead of treating death as censored in our main analysis, we treated it as a competing risk and conducted an additional sensitivity analysis using Fine-Gray subdistribution hazards regression model [42]. Third, given the exclusion of a large number of participants with missing information, missing values of the baseline covariates were imputed with multivariate imputation via chained equation (MICE) (the number of imputations was 5) to make sure that the representativeness of the cohort was not substantially affected by exclusion of participants [18].

Table 1 presents baseline characteristics of the 266,280 participants with the mean age of 55.93 [standard deviation (SD): 8.08] years. The mean duration of follow-up was 10.64 (SD: 1.34) years. There were more females (51.15%) than males and over 95% participants were of White ethnicity. About 80% participants were classified as having moderate to vigorous PA levels. The mean annual average PM_2.5 concentration in 2010 was higher in COPD incident cases [10.08 (SD 1.08) μg/m³] than in non-cases [9.93 (SD 1.04) μg/m³]. Detailed baseline information of the subsample with objectively measured PA was presented in Additional file 1: Table S2. The distribution of baseline characteristics was generally similar between the whole study population and the subsample.

Table 2 shows the associations between self-reported PA and incident COPD. Higher PA level was associated with a lower incidence of COPD. The associations were robust to adjustment for various covariates, including PM_2.5. In Model 4, compared with low PA level, the HR (95% CI) of developing COPD was 0.769 (95% CI: 0.720, 0.820) and 0.726 (95% CI: 0.679, 0.776) for moderate and high level of PA, respectively. In subgroup analysis, the negative association of self-reported PA and COPD incidence remained stable in participants exposed to different PM_2.5 levels (Table 3).

In contrast to PA, positive associations of PM_2.5 with COPD incidence were observed, and the associations remained unchanged after adjustment for PA and other covariates (Table 2). An IQR increment in PM_2.5 was associated with a 6.5% (95% CI: 3.2%, 9.9%) increase in COPD risk. In tertile-based analysis, compared with participants exposed to low level of PM_2.5, the HR of having COPD was 1.050 (95% CI: 0.987, 1.117) and 1.068 (95% CI: 1.032, 1.140) for those exposed to moderate and high levels of PM_2.5, respectively. The concentration–response curve for the association between PM_2.5 and COPD is presented in Additional file 1: Figure S1. The positive associations remained in subgroups with different PA levels, though some were statistically non-significant (Table 4).

Table 5 shows the combined health impact of self-reported PA level and chronic exposure to PM_2.5 on COPD incidence. Participants with higher PA or lower PM_2.5 exposure generally had lower risk of COPD. Using the participants with low level of PA and high PM_2.5 exposure as the reference, those with high PA level and low PM_2.5 exposure had the lowest risk of developing COPD (HR: 0.681, 95% CI: 0.607, 0.764).

Results of interaction between self-reported PA and PM_2.5 on the incidence of COPD are presented in Additional file 1: Table S3, Table S4. On additive scale, little evidence of interaction was found (Additional file 1: Table S3). Similar results were also observed on multiplicative scale, with all p values of the interaction term > 0.05 (Additional file 1: Table S4).

Overall, similar results were found in analyses among the subsample with objectively measured PA. Higher level of objectively measured PA was associated with lower risk of COPD incidence (Table 2), and the negative association remained across different PM_2.5 levels (Table 3). Participants with high PA level and low PM_2.5 exposure had the lowest risk of developing COPD (HR: 0.574, 95% CI: 0.432, 0.764) with reference to those with low level of PA and high level of PM_2.5 exposure (Table 5). No interaction between objectively measured PA and PM_2.5 were found, either on additive or multiplicative scale (Additional file 1: Table S3, Table S4).

The results of sensitivity analyses are presented in Additional file 1: Table S3-S9. Additional adjustment for baseline lung function or inverse distance to main road did not alter the results materially, especially for PA (Additional file 1: Table S3-S6). The associations between PA, long-term PM_2.5 exposure, and incident COPD also remained robust when treating all-cause death as a competing risk (Additional file 1: Table S7). Furthermore, after the imputation of baseline missing data, the effect estimates of the associations between PA and long-term PM_2.5 exposure and COPD incidence were similar as well (Additional file 1: Table S8, Table S9).

Our study has several important strengths. First, we used a large cohort of 266,280 participants with extensive information on a wide range of potential confounders, which enabled us to investigate the associations more reliably. More importantly, in addition to self-reported PA, we did an additional analysis in a subsample of nearly 100,000 participants who wore Axivity AX3 triaxial accelerometers to measure their daily PA duration and volume [24], and similar results were found. Such objective method provided more accurate measurements, which is a main challenge in studies using self-reported PA [13].

Our study also has some limitations. First, we only used the annual average PM_2.5 concentration in 2010 as a proxy for the long-term PM_2.5 exposure, which may have led to exposure misclassification. However, many previous studies suggested that the spatial distribution of PM_2.5 generally remains stable in the same region over a period as long as > 10 years [27, 29, 63]. In the UK, the annual average concentration of PM_2.5 between 2010 and 2019 were relatively stable according to the Department for Environment Food & Rural Affairs of the UK (https://​www.​gov.​uk/​government/​statistics/​emissions-of-air-pollutants/​emissions-of-air-pollutants-in-the-uk-summary). Therefore, the annual concentration of PM_2.5 in 2010 could serve as a surrogate measure of long-term exposure over the study period. Despite the issue mentioned above, another potential limitation of the study is the exclusion of a relatively large number of participants due to missing data for covariates, which generated a relatively healthier cohort [participants in the analytical cohort were more likely to be younger and physically active (Additional file 1: Table S10)]. However, the difference was not substantial, especially the difference in PM_2.5 (as small as 0.13 μg/m³), and it should not seriously bias our results. Furthermore, as an alternative approach, we additionally conducted multiple imputations for those missing covariates and the associations of PA and long-term PM_2.5 exposure with COPD incidence remained stable (Additional file 1: Table S8 and Table S9). In addition, compared with the general population, the COPD prevalence was lower in the UK Biobank cohort [the UK Biobank: 0.4% for men, and 0.4% for women; the UK general population: 3% for men, and 2% for women among participants between 55 and 64 years], which implies the evidence of a healthy volunteer bias [64]. However, it has been suggested that this may not influence the valid estimates of associations [6] since sufficiently large numbers of individuals with different levels of exposures were investigated with high internal validity [64], but may only affect the extrapolation and underestimate the associations of PM_2.5 and COPD incidence in a general population [6]. Last, our study is based on a European cohort with a lower PM_2.5 concentration [65], findings of our study may therefore not be generalizable to populations in areas with relatively high PM_2.5 concentrations.