Effects of ambient temperature on atopic dermatitis and attributable health burden: a 6-year time-series study in Chengdu, China

Methods

Study site and meteorological data collection

This study was approved by the Ethics Committee of the West China Hospital of Sichuan University (2021 Review No. 1639). The study is confined to a single geographic region—Chengdu, the capital city of Sichuan Province and the largest city in western China, with a resident population of approximately 16 million. Chengdu is in a basin area with a subtropical monsoon climate. Its annual average temperature was 16.5 °C between 2015 and 2020, with an average of 24.9 °C in the summer and 7.0 °C in the winter.

The time frame of this study spans from January 1, 2015 to December 31, 2020. Meteorological data (including daily mean temperature, relative humidity, sunshine duration, atmospheric pressure and precipitation) during this period in Chengdu were obtained from the National Meteorological Data Sharing Platform (http://data.cma.cn/). Air quality data (including PM_2.5, NO₂ and SO₂) during this period were obtained from the National Urban Air Quality Real-time Publishing Platform (http://www.cnemc.cn/sssj/). These pollution and meteorological covariates were chosen as previous studies have reported that these factors can be important confounders in identifying the real association between temperature and AD (Guo et al., 2019; Kantor & Silverberg, 2017; Wang et al., 2021). In addition, we also discussed with clinical experts and meteorologists to select the possible covariates. Approximately 2.7% of the meteorological data contained missing values, which were imputed by the linear interpolation approach, as commonly used in meteorology (Zhang et al., 2019).

Epidemiological and clinical data collection

All patients were outpatients diagnosed with AD during the same period described above in the Department of Dermatology, West China Hospital of Sichuan University, Chengdu, China. As the largest hospital in western China with more than 4,000 beds, West China Hospital ranks among the very top hospitals in China with high performance in skin disease management. From all these patients, we collected information on dates for outpatients diagnosed with AD, demographic characteristics, clinical diagnosis, current address, and related medical expenditures.

The AD outpatients were included by both searching the keywords (“atopic dermatitis” and “atopic eczema”) in the Outpatient Information System. The AD cases in our study included L20 among the ICD-10 code. Patients of all ages were included with no gender restriction and were excluded if they did not physically reside in Chengdu throughout the study period. The sample size is sufficient to ensure the statistical power in the study (R code for estimating statistical power in the Supplemental Information).

Statistical analysis

This is an ecological study with a time-series approach. We first constructed a distributed lag nonlinear model (DLNM) to estimate the effects of ambient air temperature on AD, and determine the temperature thresholds that cause an increased risk of AD. As the probability for hospital visits for AD is small, a quasi-Poisson regression was applied, allowing for over-dispersion with the distributed lag nonlinear model (DLNM). This model is widely used in environmental and health studies, due to its ability to simultaneously capture and assess the nonlinear exposure-response relationships and lagged effects of environmental factors on health outcomes (Bhaskaran et al., 2013; Gasparrini, Armstrong & Kenward, 2010). The model is expressed as follows: ${\displaystyle \text{Log}\left[E\left(Y_t\right)\right]=\alpha +\beta Tem{p}_{t,l}+NS\left(R{H}_t,3\right)+NS\left(S{H}_t,3\right)+NS\left(S{P}_t,3\right)+NS\left(T{P}_t,3\right)+\theta P{M}_{2.5}+NS\left(Time,7\ast 6\right)+\gamma DOW}$ where Y_t denotes the daily counts of outpatient visits at day t, and $E\left(Y_t\right)$ is the expected number of outpatient visits; Temp_t,l is a cross-basis matrix assessing the nonlinear and lag effects of air temperature over the current day (lag0) to the maximum lag at l days (lag l). We found that the lag effect of temperature was statistically significant in 7 days, and consequently 7 days was chosen as the maximum lag. In addition, the model had the best AIC (Akaike Information Criterion) when the lag days was set for 7 days. We used natural cubic spline function for considering the non-linear effects of temperature, and 4 degrees of freedom (df) was set based on the least AIC.

NS denotes the natural cubic spline function. $NS\left(R{H}_t,3\right)$, $NS\left(S{H}_t,3\right)$, $NS\left(S{P}_t,3\right)$ and $NS\left(T{P}_t,3\right)$ were used to control the potential confounding effects of relative humidity, sunshine duration, surface air pressure and total precipitation, respectively. They were set with natural cubic splines of 3 degrees of freedom (df) based on AIC for quasi-Poisson models. PM_2.5 represents the daily 24-hour concentrations (µg/m³) of particulate matter with a diameter of fewer than 2.5 micrometers. As different air pollutants are highly correlated, only PM_2.5was included since it was the most important contributor to air pollution in Chengdu. $NS\left(Time,7\ast 6\right)$ is a natural cubic spline of time with 7 degrees of freedom per year (all 6 years of the study) to control the seasonal and long-term trends. Based on previous methodological and epidemiological research, 7 degrees of freedom per year is sufficient to adjust for seasonal and long-term trends (Gasparrini, Armstrong & Kenward, 2010; Meng et al., 2021; Wang et al., 2021; Zhao et al., 2021). DOW denotes the day of the week and was used for controlling with-in-week variation effects. The parameter α is the intercept. θ is the coefficient parameter of PM2.5. β is the estimated coefficient, which was finally converted to the risk ratio (RR) to express the effects of temperature on AD.

Following completion of the analysis above, the optimal temperature range for preventing AD was determined based on the temperatures where the effects on the number of AD patients were not statistically significant. Additionally, subgroup analysis was conducted to identify the populations most vulnerable to temperature change. In this study, children were defined as age ≤ 18 years and adults as age >18 years. Specifically, we first divided the raw data into four different subgroups according to sex (male and female) and age (children and adults). Data from each subgroup included the daily number of AD visits, daily temperature, and other covariates. Then, we analyzed the exposure-response associations between temperature and AD in different subgroups using a unified DLNM model. Finally, we compared the exposure-response curves, attributable fraction (AF) and attributable number (AN) in different subgroups and identified individuals who are more susceptible to extreme temperatures.

In addition, we calculated the temperature-related attributable burden for AD according to the method described in previous literature (Gasparrini & Leone, 2014). The AFs and ANs were used to represent the percentage and number of AD patients affected by extreme temperature. The corresponding empirical confidence intervals (eCIs) were estimated with Monte Carlo simulations (Zhao et al., 2019). The attributable economic burden was estimated using the AN multiplied by the average health care expenditures for each patient.

Sensitivity analysis

Sensitivity analysis was performed to test the robustness of the results. For relative air humidity, sunshine duration, atmospheric pressure and precipitation, dfs in the spline function were set from 2 to 5. For the time trend, dfs in the spline function were set at 3, 5, 7 and 9 respectively. The different maximum lag times for temperature were also set at 3, 5, 7 and 9 days. We also conducted DLNM by adjusting SO2 or NO2, respectively. In addition to the AIC, we used the BIC to select the appropriate dfs and lag time, as the BIC penalizes the number of parameters in the model to a greater extent than AIC.

All statistical analyses were performed using R software (version 3.6.3, https://www.R-project.org) with the R package ‘dlnm’ (version 2.3.2). A two-tailed P value <0.05 was considered statistically significant.

Descriptive statistics

There were 10,747 outpatient visits from AD patients between 2015 and 2020 in our hospital, including 71.1% of children’s visits and 57.0% of male patient visits. Table 1 illustrates the descriptive statistics of AD outpatient visits, meteorological indicators and air pollutant concentrations. The average number of daily visits was 5.0, with 3.6 for children and 2.9 for male patients. The mean daily temperature, relative humidity, sunshine duration, surface air pressure, total precipitation, PM2.5, NO2, and SO2 were 16.5 °C, 81.0%, 2.7 h, 92.6kpa, 3.0 mm, 48.1µg/m³, 41.2µg/m³, and 11.0µg/m³, respectively. The correlations among meteorological factors and air pollution are shown in Table S1.

Figure 1 presents the distribution of the monthly average daily mean temperature and monthly mean number of AD outpatient visits between 2015 and 2020. The monthly average daily mean temperature ranged from 6.1  °C to 25.4 °C, with an overall mean of 16.4 °C during the 6-year period. The peak of AD outpatient visits occurred in the three coldest months (from November to January), followed by a lower peak in August, the hottest month. Scatter plots between AD and other meteorological indicators/air pollutants are in the Appendix (Figs. S1–S8).

Association between temperature and the number of AD outpatient visits

Figure 2 shows the overall cumulative effects of temperature on the number of outpatient visits in all AD patients. In general, the exposure-response curve followed a shallow U-shape, indicating that both high and low temperatures increased the risk of AD. Significantly strong effects were observed at temperatures below 19.6 °C and above 25.3 °C. The effect was stronger at cold temperatures than at hot temperatures. When the temperature dropped below zero the risk of AD almost doubled (RR = 2.6, 95% eCI [1.4–5.0]) compared with the centering temperature of 22.8 °C (same as the minimum mortality temperature described later). In addition, low-temperature effects lasted for 7–8 days, and high-temperature effects lasted for 3–4 days (Figs. S9–S10).

Populations vulnerable to extreme temperatures

Figure 3 shows the association between daily mean temperature and the number of AD outpatient visits in different patient groups. The exposure-response curve for the child group showed a U-shape similar to the curve for all patients in Fig. 2, suggesting that both high temperature and low temperature increased the risk of AD. However, this effect was not statistically significant in the adult group. When the patients were grouped by sex, low temperature increased the risk of AD in both the male and female groups, while high temperature increased the risk only in the female group.

Health burden of AD attributed to extreme temperatures

The health burden of AD outpatient visits attributed to extreme temperatures is presented in Table 2. Overall, 25.4% (AF = 25.4%; 95% eCI: 9.2%, 36.0%) of the AD outpatient visits were attributed to extreme temperatures, corresponding to approximately 2724 visits (AN = 2724.4, 95% eCI: 955.7, 3909.2) during the study period. Among the 10,747 visits, 22.4% were caused by low temperature and 3.0% by high temperature, leading to an estimated health care expenditures of 932698.3 RMB (137,161.5 US dollars) for AD patients in the study period.

Regarding the attributable burden for different patient groups, an estimated 29.3% (AF = 29.3%, 95% eCI: 11.5%, 40.6%) of the AD outpatient visits among children were associated with extreme temperatures, with 25.8% of them (AF = 25.8%, 95% eCI: 8.9%, 38.0%) being associated with low temperature. For male patients, an estimated 28.7% (AF = 28.7%, 95% eCI: 9.1%, 41.3%) of the AD outpatient visits were attributed to low temperature, whereas for females, 4.7% (AF = 4.7%, 95% eCI: 0.1%, 8.4%) of the AD outpatient visits were attributed to high temperature (Table 3).

Sensitivity analysis

In the above model, after changing the dfs of the spline functions for relative humidity, sunshine duration, atmospheric pressure, precipitation and time trend, and different numbers of maximum lag days for temperature, the total AFs remained statistically significant and consistent with the results described above, which indicated the robustness of the analysis (Tables S2–S3). We adjusted other covariates (SO₂ or NO₂), and the main results did not change significantly (Table S4). In addition, the result of the model based on the optimal BIC was very similar to the model based on the optimal AIC, which led to no change in the model selection.