Extreme ambient temperatures and cardiorespiratory emergency room visits: assessing risk by comorbid health conditions in a time series study

Study population and health outcome data

The study population included residents of Toronto who were admitted to an emergency department within the city from April 1^st 2002 to March 31^st 2010 with main diagnoses of cardiovascular (International Classification of Diseases, 10^th Revision [ICD-10] codes: I00-I99) or respiratory diseases (ICD-10 codes: J00-J99). The daily numbers of ER visits were identified through the National Ambulatory Care Reporting System (NACRS) [15] using the first three digits of available postal codes corresponding to the city of Toronto [15]. The NACRS database is estimated to capture more than 97% of the emergency department visits in Toronto and has been demonstrated to capture accurately the coding of the most responsible diagnosis when being admitted to the ER [16]. We also tabulated the daily number of visits for influenza (ICD-10: J09, J10, J11) in order to adjust for the potential confounding influence of viral respiratory seasonal epidemics in the analyses using respiratory disease ER visits as an outcome variable. Access to the NACRS database was granted through data sharing agreement between the Canadian Institute for Health Information (CIHI) and the Public Health Agency of Canada (PHAC).

Comorbid health conditions

The comorbid health conditions for each ER visit were defined by using all secondary ICD-10 diagnosis codes listed for the same visit as the main diagnosis of cardiovascular or respiratory diseases. The comorbid health conditions that were extracted from the secondary diagnoses are defined as follows: diabetes (ICD-10 codes: E10-E14), chronic respiratory diseases (ICD-10 codes: J40-J47), upper respiratory infections (ICD-10 codes: J00-J06), hypertension (ICD-10 code: I10), cancer (ICD-10 codes: C00-D49), kidney diseases (ICD-10 codes: N17-N19), pneumonia (ICD-10 codes: J12-J18) and cardiac diseases (ICD-10 codes: I01, I02.0, I05–I09, I11, I13, I20–I25, I27 and I30–I52). Secondary diagnoses of chronic respiratory diseases and cardiac diseases were validated by linking the data with the Hospital Morbidity Database (HMD) which captures all hospitalization in Canada and with the NACRS database. One physician service one year prior to the ER visit in either the HMD or NACRS for the medical condition under consideration was considered for identifying the comorbid conditions [8]. A secondary diagnosis of cancer was validated by linking the data with the Canadian Cancer Registry which captures all diagnoses of cancer across Canada. A one year time window before the date of ER visit was used to identify subjects with cancer. Each case of cardiovascular or respiratory disease ER visits was categorized in subgroups in a time series dataset.

Weather and air pollution data

Daily weather data were obtained from Environment Canada using the monitoring station at Toronto Pearson International Airport (latitude: 43°40′36″N; longitude: 79°37′50″W) located approximately 26 km west of downtown Toronto. Daily averages of temperature and relative humidity were computed based on hourly measurements.

Daily ambient concentrations of air pollution were obtained from the National Air Pollution Surveillance Network (NAPS) maintained by Environment Canada which collects outdoor air pollution levels through several automated fixed-site monitoring stations [17]. Daily average concentrations for the following pollutants were obtained: nitrogen dioxide (NO₂), sulphur dioxide (SO₂), carbon monoxide (CO), ozone (O₃) and particulate matter of median aerodynamic diameter less than 2.5 microns (PM_2.5). Daily values were computed by averaging across all monitoring stations the daily mean concentrations of each pollutant for each day. A minimal amount of missing values was observed for O3 and CO. We interpolated the missing values for these pollutants by using the mean of previous and the following days’ concentration.

Statistical analysis

The association between extreme ambient temperatures and daily numbers of cardiovascular ER visits or respiratory ER visits according to subgroups of comorbid health conditions was assessed using a time series study design. A quasi-likelihood Poisson regression in a generalized linear model (GLM) [18] was fitted to model the effect of daily mean temperature on the daily numbers of cardiovascular and respiratory ER visits by potentially sensitive subgroups. Daily mean temperature was chosen as the main exposure metric because it represents the exposure throughout the whole day and can be easily used for decision making purposes [19, 20]. In addition, a distributed lag non-linear model (DLNM) [21, 22] was incorporated in the Poisson regression in order to explore the non-linear relationships between temperature and daily ER visit counts cumulated across specific lag periods. We investigated effect modification by comorbid health conditions by calculating the relative effect modification (REM) index which is the ratio between the relative risk (RR) when comorbidity is present and when the comorbid health condition is absent (reference category) [23]. The REM can be interpreted as the relative increase in risk for an emergency room visit for persons with a comorbid health condition compared to those without the health condition. The statistical significance of the REM was tested with the calculation of its 95% confidence interval.

We adjusted for seasonal effects and long-term trends using a natural cubic spline for time with 10 degrees of freedom (df) per year and we accounted for sub-seasonal cycles by including an indicator variable for day-of-the-week. Several other smoothers for time were also assessed (e.g. 5, 7, 9, 11 and 13 df), but the optimal fit was suggested to be 10 df per year. A natural cubic spline was used in the DLNM with 5 df for the non-linear temperature effects and 5 df for the lagged effect. Spline knots were placed at equal spaces across the range of daily mean temperature as well as at logarithmically equal intervals in the lag space. A 21 day lag period was used to examine the effect of temperature on ER visits [24]. We also investigated adjustment for 0 to 2 lag days of moving average (i.e. average of the concurrent day and the two previous days) of specific air pollutants (NO₂, SO₂, CO, O₃ and PM_2.5) as well as the concurrent day’s average relative humidity [25]. Finally, we assessed confounding for the daily count of influenza ER visits only where the daily number of ER visits for respiratory disease was used as the main outcome variable. Evaluation of confounding for air pollutants, influenza and relative humidity was done by a backward deletion approach [26]. This was done by adjusting for all potential confounders and then by removing one by one in a stepwise manner the least significant confounding variables as long as the total proportional change in effect estimates compared with the fully adjusted model was less than 10%. Covariates that were not confounders, but increased the precision of the estimates were kept in the final model.

Effect estimates were expressed as relative risks (RRs) with 95% confidence intervals (CIs) that can be interpreted as the relative increase in daily ER visits at a specific temperature compared to a temperature reference point for persons with the medical condition under consideration. The Akaike's Information Criterion for quasi-Poisson (Q-AIC) was used to verify the optimal df of the natural cubic spline for seasonal effects, non-linear temperature effects and lagged effects [27]. All the analyses were performed using R software (version 3.0.1) with the DLNM package [22].

The effect of extreme hot temperature was evaluated by comparing the relative increase in daily ER visits of the 99^th percentile of the daily mean temperature distribution over the study period to the 75^th percentile. Extreme cold temperature effect was examined by evaluating the RR of daily ER visits associated with the 1^st percentile of temperature relative to the 25^th percentile of temperature. These cutpoint comparisons were chosen based on a previous Canadian investigation [25] in order to compare extreme temperature values to normal warm and cold temperatures. Relative increase in daily ER visits were evaluated over cumulative lag periods (i.e. 0 to 1, 0 to 13) in order to capture the possible overall short term and delayed effects of extreme temperatures on adverse health outcomes [28].

Ethics approval for this study has been obtained through a data sharing agreement between the Public Health Agency of Canada and the Canadian Institute for Health Information.