A data-driven approach to setting trigger temperatures for heat health emergencies.

Henderson, Sarah B. ; Kosatsky, Tom

Excessive mortality during unusually hot weather has been described for several cities worldwide, including Moscow in 2010, (1) Paris in 2003, (2) and Chicago in 1995. (3) Many cities have developed heat health warning systems to trigger emergency responses and to inform their populations about upcoming episodes. (4) Most systems use a two-stage approach, with lower threshold conditions triggering a "heat health advisory" and higher threshold conditions triggering a "heat health emergency". (4) One challenge in developing heat health warning systems is establishing the threshold conditions under which advisories and emergencies are triggered. Hajat et al. (5) describe the two principal approaches of synoptic classification and epidemiologic evaluation. In brief, synoptic classification combines multiple meteorological measurements (temperature, humidity, air pressure, etc.) to identify the air masses most associated with excessive local mortality. (4) Epidemiologic evaluation assumes that mortality is a smooth function of observed temperatures, and uses time series models to quantify the local effects. Regardless of the underlying methods, most heat health warning systems are developed using observed temperature data and implemented using forecast temperature data, with little consideration of the relationship between them. This has been highlighted as a limitation. (5)

Here we describe a different, more data-driven approach used to identify the trigger temperatures for heat health emergencies in Greater Vancouver, Canada. An unprecedented period of extreme hot weather resulted in excessive mortality during the summer of 2009 (Figure 1), (6,7) and local municipalities developed comprehensive heat health emergency action plans to mitigate the effects of future events. The plans call for actions such as modifying or cancelling outdoor public gatherings, allowing free access to public pools, opening cooling centres, and asking management of air-conditioned buildings (malls, theatres, etc.) to maintain longer hours. Much of this planning was done in collaboration with local public health authorities, and in early 2010 the British Columbia Centre for Disease Control (BCCDC) was asked to provide a rapid, evidence-based recommendation for emergency trigger conditions to be used that summer. Given the human and financial resources necessary to implement the action plans, the stakeholders made it clear that tolerance for false positive events (i.e., calling a heat health emergency during weather that was not unusually hot) would be low.

[FIGURE 1 OMITTED]

METHODS

Data sources

Daily all-age, all-cause mortality counts in Greater Vancouver were obtained from the BC Vital Statistics Agency for 2005 through 2009. Daily temperatures were downloaded from Environment Canada for the Vancouver (coastal, 10 km from the city centre) and Abbotsford (inland, 70 km from the city centre) International Airports during the same period. Historically-issued forecasts for Vancouver and Abbotsford were obtained from Environment Canada for the years of 2005 through 2009. Temperature forecasts for today and tomorrow are made by Environment Canada meteorologists using various computer models and output as well as conceptual knowledge of local weather. For metropolitan-area Vancouver in summertime, meteorologists consider whether high temperatures forecast across the city vary by [greater than or equal to] 4[degrees]C from the waterfront to the inland suburbs. If so, they assign a range in the forecast (near the water and inland), and we used "near water" highs whenever this was done. If not, they assign a single, regional high based on an average of temperatures across the region. A single high value is usually assigned for Abbotsford, where temperatures are less variable [Personal Communication. Lundquist D, Senior Meteorologist, Environment Canada, Kelowna, BC. Telephone conversation (October 21, 2011) with Henderson re temperature forecasting process for Greater Vancouver, BC]. During the summer months (June through August), forecasts were issued three times daily at approximately 05:00, 11:00 and 16:00. The 05:00 and 11:00 forecasts included high and low temperatures for today and tomorrow, while the 16:00 forecast included high and low temperatures for tomorrow only. Average temperatures were not forecast.

Identification of candidate triggers

Candidate triggers were identified by examining the coincidence of extreme regional mortality days (>99th percentile of all mortality between June and August) with extreme temperature days (>99th percentile of all temperatures between June and August, evaluated separately for each airport). The dates on which they coincided were defined as historical heat health emergencies, and the lowest temperature at which they coincided was identified as the candidate trigger for each airport (rounded down to the nearest degree). Analyses were conducted using the two-day average of maximum temperatures, and were restricted to the summers of 2005 through 2009 to ensure that we used the most current data over a period of relative demographic stability. Although hot-weather mortality was associated with low and high temperatures (Figure 1), we used maximum temperatures because the correlation between observed and forecast highs was stronger than the correlation between observed and forecast lows.

[FIGURE 2 OMITTED]

Predictability of historical emergencies

Days with coincident extreme temperatures and extreme mortality were defined as historical heat health emergencies. We attempted to retrospectively predict those dates using observed and forecast temperatures in the following early warning scenarios (for an emergency response starting at the beginning of the business day tomorrow):

* 28-hour lead time: The average of today's 05:00 forecast for today's high and today's 05:00 forecast for tomorrow's high.

* 22-hour lead time: The average of today's 11:00 forecast for today's high and today's 11:00 forecast for tomorrow's high.

* 19-hour lead time: The average of today's observed 14:00 temperature and today's 11:00 forecast for tomorrow's high.

* 16-hour lead time: The average of today's observed high and today's 16:00 forecast for tomorrow's high (which usually occurs at 16:00 or 17:00).

Historical heat health emergencies predicted by these scenarios were classified as true positive (heat health emergency predicted when one occurred) and false positive (heat health emergency predicted when one did not occur).

RESULTS

Candidate triggers

For the summer months in the period 2005 to 2009, daily regional all-age, all-cause mortality in Greater Vancouver (population ~2.5 million) ranged from 25 to 75 deaths, with a mean (SD) of 40 (6.8) deaths and a 99th percentile of 58 deaths. Over the same time period, the two-day average of maximum observed temperatures at the Vancouver (coastal) airport ranged from 13.0[degrees]C to 34.2[degrees]C, with a mean (SD) of 21.6[degrees]C (3.0[degrees]C) and a 99th percentile of 28.5[degrees]C. The two-day average of maximum observed temperatures at Abbotsford (inland) airport ranged from 12.5[degrees]C to 37.1[degrees]C, with a mean (SD) of 23.5[degrees]C (4.4[degrees]C) and a 99th percentile of 34.2[degrees]C. Candidate triggers for Vancouver and Abbotsford, respectively, were a two-day average of maximum temperatures [greater than or equal to] 31[degrees]C and [greater than or equal to] 36[degrees]C (Figure 2).

Prediction of historical emergencies

Historical heat health emergency dates based on the Vancouver trigger were July 29 through 31, 2009. Historical heat health emergency dates based on the Abbotsford trigger were July 11, 2007 and July 29 through 30, 2009. Thus, we were attempting to predict two different (but overlapping) sets of dates for each candidate trigger (Table 1). All historical heat health emergencies were accurately predicted in four out of twelve early warning scenarios for the Vancouver trigger and five out of twelve scenarios for the Abbotsford trigger (Table 2). There were more false positives for the Abbotsford trigger because forecast high temperatures sometimes overestimated the high temperatures observed at Abbottsford airport. On the other hand, high temperatures forecast for coastal Vancouver systematically underestimated the high temperatures observed at Vancouver airport. The minimum number of false positives for the Abbotsford trigger was two, predicted for July 22, 2006 (27 deaths, 35.8[degrees]C) and July 12, 2007 (47 deaths, 32.8[degrees]C). There were no false positives for the Vancouver trigger in three of four scenarios that correctly identified the three historical heat health emergencies.

DISCUSSION

Based on these analyses, the BCCDC recommended that a heat health emergency should be triggered for Greater Vancouver tomorrow when: 1) the average of today's 14:00 observed temperature at Vancouver International Airport and today's 11:00 forecast for tomorrow's high in coastal Vancouver is [greater than or equal to] 29[degrees]C, and/or 2) the average of today's 14:00 observed temperature at Abbotsford International Airport and today's 11:00 forecast for tomorrow's inland high is [greater than or equal to] 34[degrees]C. These were reliable indicators of the two-day average of maximum observed temperatures actually being [greater than or equal to] 31[degrees]C at the Vancouver airport or [greater than or equal to] 36[degrees]C at the Abbotsford airport (i.e., the candidate trigger conditions). This 19-hour lead-time scenario predicted all four historical heat health emergencies (Table 1) while minimizing the number of false positives based on the Abbotsford trigger. The two-day average of maximum observed temperatures on the first false positive date (July 22, 2006) was 35.8[degrees]C, which is close to the candidate trigger condition. The second false positive (July 12, 2007) would have unnecessarily extended one historical heat health emergency into a second day. This characterization of false positive triggers was important to stakeholders, who felt that emergency responders would be frustrated if multiple emergencies were triggered under temperature conditions that were perceived as unlikely to cause excess mortality.

The decision to define the 99th percentiles as extreme was based on previous work. (8) We repeated all analyses using the 95th and 97th temperature and mortality percentiles to examine the sensitivity of our results to this decision. At the 95th percentiles, the candidate trigger temperatures for Vancouver and Abbotsford would have been 26[degrees]C and 31[degrees]C, respectively, yielding a total of nine historical heat health emergencies with an impracticable minimum of 38 false positive triggers. At the 97th percentiles, the candidate trigger temperatures would have been 26[degrees]C and 31[degrees]C, respectively, capturing one more historical heat health emergency and resulting in eight more false positive triggers. Although this error rate was unacceptable to the stakeholders in Greater Vancouver, decisions related to trigger sensitivity should be informed by city-specific conditions and needs.

It was also important to stakeholders that a heat health emergency tomorrow could be reliably predicted before the end of the business day today so that responders would have as much time as possible to mobilize. Two of the 16-hour lead-time scenarios in Vancouver and Abbotsford reliably predicted all historical heat health emergencies (Table 2), but stakeholders were adamant that they needed at least a few more hours of warning; we therefore began to explore the relationship between hourly temperatures and daily high temperatures. Although the 12:00, 13:00 and 14:00 temperatures were all strongly correlated with all daily highs (typically observed at 16:00 or 17:00), only the 14:00 temperature was strongly correlated with the daily highs on very hot days ([greater than or equal to] 32[degrees]C). This relationship allowed us to extend the lead time to 19 hours, giving responders 2-3 business hours to prepare for a heat health emergency on the following day.

We also suggested using temperature data from other regional weather stations, but stakeholders expressed a strong aversion to this option, preferring to keep the trigger simple and easy for all of its users (health authorities, municipalities and emergency responders) to understand. This echoes concerns the authors have heard about the complex synoptic classification system used in Toronto, (9) and is consistent with the recommendation that heat health warning systems "should be developed with all relevant stakeholders to ensure that the issues of greatest concern are identified and addressed, thus increasing the likelihood of success". (10) Most trigger-setting approaches are quite complex, (4,5) and there is little discussion in the descriptive literature about how that complexity affects their users. For example, the heat health watch warning established in the city of Philadelphia uses a complicated, multi-stage algorithm for identifying heat health advisory and emergency days, but the regional health commissioner (likely an individual without meteorological training) has ultimate responsibility for making final judgement calls. (11) This is not to suggest that such systems are limited because they are complex, but to highlight another strength of the simpler approach described here.

The summers of 2010 and 2011 were not unusually hot in Greater Vancouver, with no heat health emergencies triggered and no excessive mortality observed on moderately hot days. We therefore cannot evaluate the efficacy of the system since its initiation. Regardless, the BCCDC used these methods to provide rapid, data-driven, and evidence-based recommendations about trigger conditions for regional heat health emergencies, thereby enabling the implementation of already-existing emergency action plans to protect public health during any hot-weather events that might have occurred. Stakeholder engagement at all stages ensured that our approach maximized the likelihood of identifying real heat health emergencies, minimized the impact of false positive triggers, and remained clearly understood by its users. The use of forecast temperatures in the development stage ensured that the triggers were tested under the most realistic conditions. The Greater Vancouver system and its triggers will be continually evaluated and revised as new data become available, as has been identified as a key component of any heat health early warning system.

Acknowledgements: The British Columbia Centre for Disease Control gratefully acknowledges the guidance and contributions of our partners at the Vancouver Coastal Health Authority, Fraser Health Authority, Environment Canada, and Health Canada.

Conflict of Interest: None to declare.

Received: February 14, 2012

Accepted: April 21, 2012

REFERENCES

(1.) Osborn A. Moscow smog and nationwide heat wave claim thousands of lives. BMJ2010;341:c4360.

(2.) Fouillet A, Rey G, Laurent F, Pavillon G, Bellec S, Guihenneuc-Jouyaux C, et al. Excess mortality related to the August 2003 heat wave in France. Int Arch Occup Environ Health 2006;80(1):16-24.

(3.) Whitman S, Good G, Donoghue E, Benbow N, Shou W, Mou S. Mortality in Chicago attributed to the July 1995 heat wave. Am J Public Health 1997;87(9):1515-18.

(4.) Sheridan SC, Kalkstein LS. Progress in heat watch-warning system technology. Bull Am Meteorological Soc 2004;85(12):1931-41.

(5.) Hajat S, Sheridan SC, Allen MJ, Pascal M, Laaidi K, Yagouti A, et al. Heat-health warning systems: A comparison of the predictive capacity of different approaches to identifying dangerously hot days. Am J Public Health 2010;100(6):1137-44.

(6.) Kosatsky T, Henderson SB, Pollock SL. Shifts in mortality during a hot weather event in Vancouver, Canada: Rapid assessment with case-only analysis. Am J Public Health In press.

(7.) Kosatsky T. Hot day deaths, summer 2009: What happened and how to prevent a recurrence. BC Med J2010;52(5):261.

(8.) Medina-Ramon M, Zanobetti A, Cavanagh DP, Schwartz J. Extreme temperatures and mortality: Assessing effect modification by personal characteristics and specific cause of death in a multi-city case-only analysis. Environ Health Perspect 2006;114(9):1331-36.

(9.) Kosatsky T, King N, Henry B. How Toronto and Montreal (Canada) respond to heat. In: Kirch W, Bertollini R, Menne B (Eds.), Extreme Weather Events and Public Health Responses. Heidelberg, Germany: Springer Berlin, 2005;167-71.

(10.) Ebi KL. Towards an early warning system for heat events. J Risk Res 2007;10(5):729-44.

(11.) Kalkstein LS, Jamason PF, Greene JS, Libby J, Robinson L. The Philadelphia hot weather-health watch/warning system: Development and application, summer 1995. Bull Am Meteorological Soc 1996;77(7):1519-28.

Sarah B. Henderson, PhD, Tom Kosatsky, MD

Authors' Affiliation

British Columbia Centre for Disease Control, Environmental Health Services, Vancouver, BC

Correspondence: Sarah Henderson, BCCDC, Environmental Health Services, 655 West 12th Avenue, Vancouver, BC V5Z 4R4, Tel: 604-707-2449, Fax: 604-707-2441, E-mail: sarah.henderson@bccdc.ca

Table 1. Summary of the Historical Heat Health Emergencies for the
Vancouver and Abbotsford Candidate Triggers

Date             Number       True Positive       2-day Average of
                of Deaths     for Vancouver     Maximum Temperatures
                            Trigger ([greater       Observed at
                            than or equal to]    Vancouver Airport
                              31[degrees]C)         ([degrees]C)

July 11, 2007      58              No                   26.2
July 29, 2009      63              Yes                  32.5
July 30, 2009      75              Yes                  34.2
July 31, 2009      61              Yes                  31.6

Date              True Positive       2-day Average of
                 for Abbotsford     Maximum Temperatures
                Trigger ([greater       Observed at
                than or equal to]    Abbotsford Airport
                  36[degrees]C)         ([degrees]C)

July 11, 2007          Yes                  36.4
July 29, 2009          Yes                  37.1
July 30, 2009          Yes                  36.8
July 31, 2009          No                   32.2

Table 2. Summary of the Prediction Results for Both Triggers Under
Different Lead-time Scenarios

                                            Vancouver
                                       (candidate trigger
                                     [greater than or equal
                                       to] 31[degrees]C)

Predicting              Lead Time   True Positive    False
                                      (max = 3)     Positive

Candidate trigger        28-hour          0            0
temperature              22-hour          0            0
                         19-hour          0            0
                         16-hour          1            0

1[degrees]C less than    28-hour          1            0
candidate trigger        22-hour          2            0
temperature              19-hour          1            0
                         16-hour          3#           0#

2[degrees]C less than    28-hour          2            1
candidate trigger        22-hour          3#           1#
temperature              19-hour          3# *         0# *
                         16-hour          3#           0#

                                      Abbotsford Combined
                                      (candidate trigger
                                     [greater than or equal
                                       to] 36[degrees]C)

Predicting              Lead Time   True Positive    False
                                      (max = 3)     Positive

Candidate trigger        28-hour          1            2
temperature              22-hour          1            1
                         19-hour          1            0
                         16-hour          1            2

1[degrees]C less than    28-hour          1            2
candidate trigger        22-hour          2            2
temperature              19-hour          1            2
                         16-hour          3#           3#

2[degrees]C less than    28-hour          3#           4#
candidate trigger        22-hour          3#           5#
temperature              19-hour          3# *         2# *
                         16-hour          3#           4#

                                            Combined

Predicting              Lead Time   True Positive    False
                                      (max = 3)     Positive

Candidate trigger        28-hour          1            2
temperature              22-hour          1            1
                         19-hour          1            0
                         16-hour          2            2

1[degrees]C less than    28-hour          2            2
candidate trigger        22-hour          3            2
temperature              19-hour          2            2
                         16-hour          4#           3#

2[degrees]C less than    28-hour          3            5
candidate trigger        22-hour          4#           6#
temperature              19-hour          4# *         2# *
                         16-hour          4#           4#

Scenarios in which all historical heat health emergencies were
identified are marked in bold. The scenarios that minimize false
positive triggers while maximizing lead time are marked with an
asterisk (*). The incidence of false positives was higher using data
for Abbotsford airport because forecast temperatures sometimes
overestimated the observed temperatures.

Note: Scenarios in which all historical heat health emergencies were
identified are marked in bold indicated with #.