Study design and population
The Guangdong Chronic Disease and Risk Factors Surveys are a series of provincially representative surveys, which were conducted by Guangdong Provincial Center for Disease Control and Prevention in 2007, 2010 and 2013 and 2015. These surveys aimed to understand the prevalence trend and risk factors of non-communicable diseases such as hypertension, T2DM, and obesity. The questionnaire survey, anthropometric measurements and laboratory analysis followed the same procedure and method. The data from the four surveys were combined to examine the association between ambient temperature and FPG. All participants agreed to participate and signed informed consents form prior to the surveys. The study was approved by the Ethics Committee of Guangdong Provincial Center for Disease Control and Prevention (Ethical review code: 2019025).
Similar sampling protocols were adopted for the surveys of 2007, 2010 and 2013, which has been described previously elsewhere [12]. Briefly, in each of the surveys, 21 districts or counties in Guangdong province were randomly selected by stratified multistage cluster sampling with probability proportional to size. In the second stage, four neighborhoods or townships from each district or county were selected; In the third stage, three communities or villages from each neighborhood or township were chosen; In the fourth stage, 50 to 100 households from each community or village were randomly sampled; Finally, 1 resident aged ≥18 years from each sampled household was selected using the Kish grid method. If there is no resident ≥18 years in the selected household or if the selected resident did not agree to participate in the survey, the household was replaced with another randomly selected household nearby. Details of sampling methods and survey protocols for the nutrition and health survey conducted in 2015 have been described in the previous study [13]. The sample size and survey site of four surveys were show in Table S1 (Additional file 1).
FPG measurement and T2DM definition
Participants were asked to fast at least 8 h before blood collection. Fasting blood samples were collected by registered nurses. FPG levels were measured on a Hitachi 7600 automatic biochemical analyzer (Hitachi, Ltd., Tokyo, Japan) using reagents obtained from Wako Pure Chemical Industries Ltd. at the National CDC of China. According to World Health Organization 2006 criteria [14], known T2DM patients were defined as physician-diagnosed T2DM (confirmed with medical history). Newly detected T2DM patients were defined a new detection of diabetes with an FPG level of 7.0 mmol/L or over among undiagnosed diabetes (without a history of diabetes and hypoglycemic use), and normal fasting glucose (NFG) participants were defined as participants with an FPG level less than 7.0 mmol/L. The T2DM prevalence was defined as the proportion of known T2DM patients and undiagnosed diabetes with an FPG level of 7.0 mmol/L or over. The glycemic control rate was defined as the proportion of known T2DM patients with FPG less than 7.0 mmol/L [14].
Data collection
Questionnaire survey and anthropometric measurements
Participants were interviewed and provided with onsite health examinations. All interviews and examinations were conducted following standard protocols by physicians who had received specific training for the survey and health examination. Questionnaires were used to collect a wide range of information including demographic characteristics, lifestyle and household location, as previous studies described [12, 13]. Demographic characteristics included age, sex, career, education. Physical activity time was defined as leisure time spend in high intensity sports or moderate intensity exercise, such as running, swimming, doing Tai Chi (in hour/day). Sedentary leisure time was defined as time spent in sedentary activities after work, such as watching TV, reading a newspaper and using a computer (in hour/day). Smoking status was measured by whether smoking in the past or present (yes vs no). Drinking status was defined as whether drinking alcohol in the past 12 months (yes vs no). Height and weight were measured following standard protocols. Body mass index (BMI) was calculated as weight divided by the square of height (in kg/m2). The information of using hypoglycemic medicine in known T2DM subgroup was also collected. In addition, the information of the weekly food consumption of grains, vegetable, fruit, meat and family history of diabetes was also collected in the surveys of 2010 and 2015.
Meteorological data
Daily meteorological data including daily mean, minimum, maximum temperature (°C), relative humidity (%) and sunlight (hour/day) during 2007–2016 of 86 weather monitoring stations were obtained from the Guangdong Meteorological Service. Daily meteorological data was passed through quality control checks. The completeness of each meteorological data was closely to 99.9%. Our survey sites (district or county) were shown in Table S1 in additional file 1. In order to obtain a more accurate measure of exposure, we used an inverse distance weighted (IDW) method to produce a 1 km × 1 km spatial resolution of daily temperatures, relative humidity and sunlight across Guangdong province. The results of 10-fold cross-validation show good prediction accuracy of the interpolation method for daily mean temperature (R2 = 0.98, RMSE = 0.82 °C), daily minimum temperature (R2 = 0.98, RMSE = 1.04 °C), daily maximum temperature (R2 = 0.98, RMSE = 1.05 °C), relative humidity (R2 = 0.82, RMSE = 5.10%), sunlight (R2 = 0.87, RMSE = 1.43 h/day) (Fig. S1 in additional file 1). Then, daily meteorological data of each participant were extracted from the corresponding interpolated grid according to their residential districts/counties. We collected the lags 0–6 day (24-h) daily temperatures at the date of survey.
Statistical analysis
We described distributions of all variables, continuous variables as the means±SD for normally distributed data and median (25th–75th percentile) for skew distributed data. Categorical variables were expressed as numbers and percentages. T-test (for normally distributed continuous data), Kruskal–Wallis test (for skew distributed continuous data) or χ2 test (for categorical variables) were used to compared the difference between NFG, known T2DM and newly detected T2DM subgroups. A Gaussian generalized additive mixed models was used to investigate the relationship between ambient temperature and FPG in different subgroups after adjusting for covariates. According to the previous study [2, 15], covariates included age, sex, BMI, education, career, physical activity, sedentary leisure times, smoking status, drinking status, humidity and use of hypoglycemic medicine. Each district or county was modelled as a random effect. Daily mean temperature was selected as exposure according to the minimum value of Akaike’ s Information criterion (AIC) in the model (see Fig. S2 in the Additional file 1). And daily mean temperature, humidity, age and BMI were fitted using a penalized cubic spline function with a degree of freedom (df) of 3. The selection of optimal df of daily mean temperature was based on graphic smoothness and minimum value of AIC (see Fig. S3 in Additional file 1). The regression model was described as the following.
$$ {Y}_{im}={\beta}_{0m}+{\beta}_{1m}\ s\left({X}_{temp},k=3\right)+{\beta}_{2m}\ s\left({X}_{humidity},k=3\right)+{\beta}_{3m}\ s\left({X}_{age},k=3\right)+{\beta}_{4m}\ s\left({X}_{BMI},k=3\right)+{\beta}_{5m}\ {X}_{1i}+\cdots +{\beta}_{nm}\ {X}_{ni}+{\upxi}_{\mathrm{j}}+{\epsilon}_{im} $$
(1)
Where m represents groups (total/ NFP / newly detected T2DM/diagnosed-T2DM participants), Yim represents participant’s FPG; β0m is the overall intercept, β1m…βnm corresponds to coefficients for covariables. Xtemp, Xhumidity, Xage, XBMI and X1i…Xni denotes covariables. S() is a penalized cubic spline function, ξj is the district/county random effect and ϵim is the residual error.
In order to check the magnitude of the association between temperature and FPG differed in subgroups (NFG participants, known T2DM patients and newly detected T2DM patients), we added an interaction term of temperature and T2DM status variable on generalized additive mixed model in total population.
After constructing the model, we obtained the curve between daily mean temperature and the difference of FPG compared to the minimum FPG temperature. In order to quantitatively estimate the association between temperature and FPG, we calculated the difference of FPG comparing the minimum/maximum temperature with the minimum FPG temperature. Based on curve, we could compute the difference of FPG (∆FPGijm) at different ambient temperature. Temperature-adjusted FPG (FPG2ijm) (2), prevalence (Rate1j) (3) and glycemic control rate (Rate2j) (4) of T2DM can be estimated as the follows:
$$ {FPG}_{2ijm}={FPG}_{1im}-\Delta {FPG}_{ijm} $$
(2)
$$ {Rate}_{1j}={N}_{1j}/{N}_3 $$
(3)
$$ {Rate}_{2j}={N}_{2j}/{N}_4 $$
(4)
Where m corresponds to groups (total/NFG/newly detected T2DM/known T2DM participants); j represents reference temperature points (5 °C, 10 °C, 15 °C, 20 °C, 22.5 °C, 25 °C and 30 °C). ∆FPGijm corresponds to the difference of FPG at temperature on the survey date compared to reference temperature points. FPG1im is each participant FPG; N1j represents the sum of the number of known and newly detected T2DM patients and the number of NFG subgroup with FPG2ijm level of 7 mmol/L or greater; N3 is the number of total population; N2j represents the number of known T2DM with a FPG2ijm less than 7 mmol/L; N4 is the number of known T2DM patients.
In sensitivity analysis, we analysis the association between ambient temperature and FPG at different lags (lag1 to lag6). We further added sunshine and precipitation to the model to test the robustness of that association. we also performed sensitivity analysis using the participants whose has weekly food consumption and family history of diabetes information. We reanalyzed the temperature-FPG relationships among subgroup of NFG and newly detected T2DM when we defined T2DM using both FPG and 2-h plasma glucose rather single FPG. We used R software (version 3.5.1, R foundation for Statistical Computing, Vienna, Austria). All statistical tests were two-sided and P values of all statistical analyses less than 0.05 was considered statistically significant. The GAM analysis was performed by using package “mgcv”.