From August 1^st to 5^th, 2003, the average maximum temperatures recorded in France increased from a value close to the normal value (25°C) to 37°C, then remained between 36 and 37°C until August 13^th, before beginning to fall (28°C on August 16^th). Almost all of the population of France, i.e. approximately 60 million people, was exposed to the heat wave: the temperature exceeded 35°C for at least 9 days in 61 of the 96 French departements. A few days after the beginning of the heat wave, information reported by the medical emergency services, fire department and hospital emergency departments showed an abrupt increase in the number of emergency interventions and in the mortality accompanying the heat wave.

The mortality associated with heat has been described and reviewed by several authors. [3,4]

Some authors have described the association between the usual daily fluctuations in temperature and mortality using time series or a spatial approach. [6,9,10,26,28,31,33,37,40,44,46,47]

Others have focused on individual major heat episodes, in which an exceptional increase in mortality was observed concomitantly with an exceptional heat wave. Those authors described the characteristics of the mortality during or after the heat wave. [1,6,9,11,12,19,25,28,35,38,39,40,42,48] Studies of the 2003 heat waves in Western Europe have been conducted for Portugal,[36] Spain,[45] Italy,[8] England and Wales,[24] the Netherlands,[13] and Switzerland.[15] The probable increase in the frequency of such events in the 21^st century has been stressed, as have the time-and place-dependence of the main characteristics of that mortality.

Given that context and using the latter approach, the epidemiological characteristics of the excess mortality associated with the heat wave in France in August 2003 have been reported herein, on the basis of the studies carried out at the French National Institute for Health and Medical Research (Inserm) by the authors and publicly reported to the French health authorities. [17,18]

The observations made from August 1^st to November 30^th, 2003 in mainland France, were compared to the similar observations for the period 2000–2002, taken as the reference period. The mortality data were derived from the merged data bases of Inserm and INSEE (National Institute of Statistics and Economic Studies), thus ensuring that the data were complete.

A prior descriptive analysis of the daily mortality-rate process over the reference period has shown that, when the seasonal effect and linear annual trend were controlled, the residuals were distributed as a first-order autoregressive process with extra-Poisson variability. Two methods were then used to estimate the expected number of deaths and the expected variance of daily death counts in 2003. Method A combined modelling of the time course of monthly mortality rates, by age and gender, from 2000 to 2002, extrapolated to 2003, with estimations of the populations by year, age and gender, from 2000 to 2003. The model adopted was a log-linear Poisson model incorporating a linear annual trend and a specific term for each month of the year. Method B consisted in estimating the mean mortality rates by age, gender and month, from 2000 to 2002, and applying them to the 2003 population estimates. In both cases, an over-dispersion parameter was estimated and applied to parameter variance estimation. Ten-year age groups were considered.

For France as a whole, method A was adopted, since the daily numbers of deaths observed from March to June 2003 were, on average, closer to the estimated values calculated using method A, did not differ significantly from them, and differed significantly from the estimated values calculated using method B. At the finer geographical scale of the 22 “regions” and 96 “departements” that constitute mainland France, method B proved superior to method A on the basis of the same criteria.

Given the observed (O) and the expected (E) numbers of deaths, mortality was quantified using excess mortality (O-E), rounded up to the nearest integer, and mortality ratios (O/E). On the basis of the fluctuations in the daily numbers of deaths that were observed during the reference period (days of August in 2000, 2001 and 2002), “95% fluctuation intervals” were defined as intervals in which there was a 95% probability of observing the daily numbers of deaths in the future if they had the same expected value, variance and auto-correlation as the deaths which were observed during the reference period. “Fluctuation intervals” are therefore prediction intervals.

The 95% fluctuation intervals of the observed mortality around the expected mortality for the period 2000–2002 were estimated by day, ten-day period and month, taking into account the autoregressive structure of the daily mortality-rate process and the extra-Poisson variability observed over the reference period.

The analysis of the causes of death was conducted on the causes of death reported by the physicians making out the death certificates as the “initial” cause. The latter is defined as the cause initiating the morbid process resulting in death. The causes of death were classified using the International Classification of Diseases (ICD10) and an additional category for deaths reported to be directly related to the heat wave by the physician, i.e.: dehydration, hyperthermia and heatstroke.

Usually, coding the medical causes of death involves a long validation process. Validation is still ongoing for 2002. In the 2003 heat wave context, validation was accelerated for August 2003. It was thus possible to compare the number of deaths by cause occurring in France in the first 20 days of August 2003 with the means of the same death counts for August 2000 and 2001, considered to be the expected number of deaths by cause. The statistical significance of the difference between the observed and expected number of deaths due to a given cause was determined under the hypothesis that the numbers of deaths used in the comparisons were independently Poisson distributed.

The daily maximum and minimum temperature statistics used are those derived from the Météo-France network of 192 meteorological stations representative of the cities of France. For each of the 96 departements of France, a “number of very hot days” was determined. The number was defined as the number of days on which the minimum and maximum temperatures reported by the departement‘s meteorological stations simultaneously exceeded the corresponding 30-year averages by 5 and 9°C, respectively, between August 1^st and 20^th, 2003. The temperature cut-off points (5 and 9°C) were chosen so as to minimize the deviance of the relationship between the mortality ratios and the number of very hot days using a Poisson regression. The 96 France departements were divided into four groups with nearly equal numbers of expected deaths and an increasing number of very hot days: 0 to 1 day; 2 to 3 days; 4 to 7 days; 8 to 13 days.

In the first fortnight of August 2003, the entire population of the 60 million inhabitants of France was exposed to a heat wave of unprecedented amplitude. The first significant excess mortality was observed 3 days after the maximum temperatures rose above the seasonal normal value. Over the 9 consecutive days that maximum temperatures remained 11 to 12°C higher than the seasonal average, the daily excess mortality increased constantly. Mortality returned to normal for the overall population of France on August 19^th, 4 days after the maximum temperatures had returned to values close to the seasonal averages. Almost 15 000 excess deaths occurred between August 1^st and 20^th, 2003, i.e. 55% more than the expected mortality.

On a larger scale, from June to August 2003, the whole of Western Europe was struck by a series of heat waves. Excess mortality was observed in Portugal (+1316 deaths),[36] Spain (+6595 to 8648 deaths),[45] Italy (in 21 cities: +3134 deaths),[8] England and Wales (+2139 deaths),[24] the Netherlands (+1400 to 2200 deaths),[13] and Switzerland (+975 deaths).[15] Therefore, the overall excess death count of the 2003 heat wave is higher if all Western European countries are considered from June to August. A large fraction of the excess deaths appears to have occurred in France between August 1^st and 20^th, with a marked increase in mortality applied to a large population (60 million people).

Numerous other episodes of major excess mortality associated with heat waves have been reported in North America,[1,6,11,19,25,35,42,48] Europe,[9,12,28,39,40] and Japan.[38] As was the case in those reported heat waves, the excess mortality observed in France in August 2003 increased with age.[1,2,37,42,48] It is, however, noteworthy that even though the majority of the excess mortality affected subjects aged 75 years and over (+11 731 deaths), very marked excesses were nonetheless observed in all age groups from 35 to 74 years (+2930 deaths). Moderate but significant excess mortality also occurred among male infants aged less than 1 year (+25 deaths). If the elderly are a particularly vulnerable population, the belief that they only are at risk of heat-related death would thus be erroneous.

Both women,[34,39] and men,[48] have been reported to have higher mortality ratios than the opposite gender. In this study, the overall excess mortality for women was 75% higher than that for men. However, most of this difference may be explained by the greater longevity of women. When equal ages are considered, from age 55 years, there is only a female excess mortality of about 15%. While the difference needs to be elucidated, it would be imprudent to consider that only women are exposed to the risk of death in heat waves.

While the relative increase in mortality observed in hospitals and clinics was less than that for subjects at home or in retirement homes, the absolute excess mortality in the public hospitals was nonetheless major. Thus, all the medical emergency services and emergency care management systems were subject to exceptionally intense demand. The particularly high mortality ratios observed among people living at home and the substantially higher mortality ratios observed for single and divorced subjects compared to married subjects point to the key role of isolation in vulnerability to heat waves.[29,35,43] People living alone at home are to be considered a sub-population exposed to a particularly high risk.

The deaths occurring during the heat waves described in the literature were related to a wide variety of diseases. [1,11,19,34,39,42,43] Among the variety of medical causes of the excess deaths in France in August 2003, some reflect the very morbid process which resulted in death. This is the case for the deaths due to heatstroke, hyperthermia and dehydration, but also, probably, for a fraction, at least, of the deaths for other diseases, particularly circulatory and respiratory diseases.[1,11,19,20,34,39,42] Certain medical causes of death may also reflect special vulnerability to a heat wave associated with a chronic disease. This was particularly the case for deaths related to mental disorders, nervous system diseases,[2,34,35,43] and endocrine diseases,[34,42] but also, probably, for other causes.

The excess mortality was substantially greater in the departements that had the greatest number of very hot days and more markedly in the older age groups. However, the relationship is far from perfect: the meteorological indicator of heat-wave exposure used herein is, in fact, only a parameter enabling classification of the 96 French departements on an ordinal scale of heat-wave exposure. An analysis of heat heterogeneity on a finer geographic scale is required. In addition, several heat-wave characteristics appear to have a strong impact on mortality: an abrupt major increase in temperature, whose impact is immediate or very short term;[26] a succession of very hot days;[14,47] and the time during the summer at which the heat wave occurs. [10] The exact form of the impact is still far from being elucidated. Heat waves may be described using multiple meteorological indicators such as minimum, maximum or apparent temperature and relative humidity.[12,31,37] The latter remained particularly low during the August 2003 heat wave and the daily meteorological data for that indicator were therefore not taken into account herein.

The excess mortality observed showed great geographic heterogeneity ranging from +20% in the least-affected coastal regions to +142% in the Paris area region. The particularly marked excess mortality in the city of Paris and the immediately suburbs is probably related to the higher risks of death in large metropolitan areas reported by certain authors. [6,28] This observation is to be considered in the context of the heat island concept,[6,7,32,42,46] and the question of whether there is a synergy between prolonged high temperatures and socioeconomic characteristics.[19,25,30,34,39,47] Some studies also suggest a possible contribution of atmospheric pollution to the increase in the number of deaths during heat waves.[26,28,33,40,44] In August 2003, ozone and particulate matter concentrations were particularly high in France: the ozone concentration alert threshold (240 (μg/m3) was exceeded for at least one recording station for 12 consecutive days from August 2^nd to 13^th, 2003. Moreover, exceptionally high values were observed in areas where the ozone concentration is generally low. [21]

The above factors could explain some of the differences between the geographic distributions of mortality ratios and the number of “very hot days”. Both a finer geographic analysis of temperatures and further investigation of the vulnerability factors related to socioeconomic characteristics are required.

Few observations relating to the existence of a harvesting effect following earlier heat waves are available. They show that a harvesting effect was limited,[5,9,16,26,40] or even non-existent.[20,39] In the present study, no harvesting effect was observed; the excess mortality of the August 2003 heat wave was not followed by a less-than-expected number of deaths until the end of 2003. Moreover, the excess mortality in the first 20 days of August 2003 was not followed, from August 21^st to November 30^th, 2003, by persistent excess mortality.

According to the IPCC’s climate change models, [23] average temperatures should increase by 1.4 to 5.8°C between 1990 and 2100. In 2100, the annual probability of a summer heat wave similar to that which struck western Europe in August 2003 could be greater than 50%.[41] Given that context, the setting up of systems to prevent the risks related to heat waves, combining long-term preventive measures with alert systems based on short term weather forecasts, needs to be seriously considered even in countries which have, so far, been relatively spared by heat waves. Several preventive and alert systems have been set up in response to heat waves associated with marked excess mortality (http://www.ci.mil.wi.us).[27,35,43] In France, the Institut de Veille Sanitaire (InVS), in cooperation with Météo-France set up such a system in June 2004. [22]

The epidemiological knowledge that could orient such preventive actions and alert systems is currently far from complete, given that the climate configuration of population exposure to heat waves and the population sensitivities thereto appear to be population-and time-dependent [3,4]. Research on modeling heat-mortality relationships and on the individual and collective vulnerability factors that modulate those relationships is therefore required.

The findings reported herein show that while there are populations that are particularly vulnerable to heat waves, such as the elderly and people living alone, no segment of the population may be considered protected from the risks associated with heat waves.