Tens of thousands additional deaths annually in cities of China between 1.5 °C and 2.0 °C warming

The increase in surface air temperature in China has been faster than the global rate, and more high temperature spells are expected to occur in future. Here we assess the annual heat-related mortality in densely populated cities of China at 1.5 °C and 2.0 °C global warming. For this, the urban population is projected under five SSPs, and 31 GCM runs as well as temperature-mortality relation curves are applied. The annual heat-related mortality is projected to increase from 32.1 per million inhabitants annually in 1986–2005 to 48.8–67.1 per million for the 1.5 °C warming and to 59.2–81.3 per million for the 2.0 °C warming, taking improved adaptation capacity into account. Without improved adaptation capacity, heat-related mortality will increase even stronger. If all 831 million urban inhabitants in China are considered, the additional warming from 1.5 °C to 2 °C will lead to more than 27.9 thousand additional heat-related deaths, annually.

C limate change is the biggest global threat of the 21st century 1 . Adverse weather events are projected to increase dramatically in frequency, severity, and duration. Global warming is projected to affect human health, with primarily negative consequences of increasing number of excess deaths and hospital admission worldwide [2][3][4][5] . In the recent past, numerous extreme high temperature events with associated mortality have taken place worldwide. For instance, the heat wave of 2003 in Europe resulted in more than 70,000 additional deaths 6,7 . An unprecedented high temperature event in Moscow and Western Russia in the summer of 2010 led to nearly 55,000 excess deaths 8,9 . A record-breaking high temperature event in Shanghai, China in 2013 brought 160 excess deaths in Pudong New District alone 10 . Considering ever worsening situation, it is of utmost importance to project the adverse health effects of high temperature to support the developing of targeted intervention strategies for public health protection.
Impacts of future climate extremes on public health have been a major research topic in recent years 5,[11][12][13][14][15] . The Special Report on Global Warming of 1.5°C emphasized that, with high confidence, an increase in heat-related mortality caused by high temperature at 1.5°C and 2.0°C threshold levels is apparent 16 . Although the decrease in cold season low-temperature extremes is expected to result in lower mortality rates during the winter months, the increase in heat-related mortality could outweigh such reductions in cold-related mortality, even in regions with colder climate 3,17,18 . Studies have consistently projected that a warmer future will lead to increases in future mortality with tens of thousands of additional premature deaths per year in the United States, and over a hundred thousand per year globally 19,20 . Still, projecting changes in future health impacts associated with climate warming remains challenging and involves large uncertainties. In particular, little is known about future impacts of heat waves in less developed countries, where capacity to address climate change is comparatively low and vulnerability to climaterelated damages is high.
Most projections of heat-related mortality under climate change did not account for population acclimatization to heat stress. People may adapt to heat stress through modifications in activities, increased use of air conditioning, and alternative building designs 21 . Projecting future mortality effects of climate change without considering human adaptability may lead to a substantial overestimation 22,23 . On the other hand, due to differences of the gender-and age-related physiological and thermoregulatory properties, increase in vulnerable population may amplify future heat-related health impacts. The fact that changes in these demographic structures have not been considered in previous studies may have caused an underestimation of mortality due to climate change 5,14,[24][25][26][27] .
China is the largest developing country, and has a faster increase in surface air temperature than the global average 20,28,29 . The elderly population is increasing and will continue to increase further in the 21st century even after the end of the one-child policy. As a result, the heat-related health risk will probably be aggravated in future. However, only a few studies focused on heat-related health impacts in China 11,14,20,27,30,31 , and they often ignored the changing population structure and adaptation capacity. In our study, the heat-related mortality in major cities of China is assessed by applying case analyses from 27 metropolises (Supplementary Fig. 1 and Supplementary Table 1) for 1.5°C and 2.0°C global warming. The mortality projections are based on an integrated assessment framework that combines projected high temperature from multiple GCMs, predicted population by gender and age structure under five SSPs, and a dynamic temperature-mortality relationship with consideration of improving adaptation capacity. In addition to the changes in the mortality-inducing high temperature, the differences of mortality between various climate and socioeconomic scenarios are also assessed to deepen our understanding of the potential benefits of climate change mitigation that will limit global warming.

Results
Definition of threshold temperature. Global mean surface air temperature of 1986-2005 was by 0.61°C warmer than the preindustrial level 32 , and further increase to 0.87°C (likely between 0.75°C and 0.99°C) for the decade 2006-2015 was reported 16  Existing studies identified a non-linear U-, V-or J-shaped relationship between temperature and mortality, suggesting that the mortality will sharply increase once a certain threshold is exceeded 5,[36][37][38][39][40] . We classified all heat-related mortality cases of 27 metropolises during the time period 2007-2013 into four groups by gender (male and female) and age (working age: 15-64 years and non-working age: ≤14 and ≥65 years). In the follow-up, a distributed lag non-linear model (DLNM) was applied to identify the temperature-mortality relationship for each group. The DLNM model is used to estimate the relative risk (RR) of mortality for each temperature, and RR = 1 corresponding to the mortality-inducing threshold temperature (see "Methods", Supplementary Fig. 3 and Supplementary Table 3). Once daily maximum temperature reaches or exceeds the threshold, these days are counted as days with high temperature. The intensity of high temperature is defined as the range of temperature (in degrees Celsius) over the threshold.
Trends in high temperature. Temperature thresholds of mortality vary for different gender and age groups. The lowest threshold corresponding to mortality-inducing temperature for female non-working age population was selected to assess the changes of frequency and intensity of high temperature in each China metropolis. According to the ensemble mean of 31 GCM outputs, annual frequency of high temperature averaged over 27 metropolises shows a significant positive trend of 1.5d/10a during 1961-2005, and continuously, a significant upward trend is projected until the 2050s. The rate of the increase will go to zero (RCP2.6) or slow down (RCP4.5) after the 2050s. With global warming of 1.5°C or 2.0°C, on average, 67.1 or 73.8 days of high (mortality-inducing) temperature, respectively, will occur per year in 2060-2099. This is an increase by 32.6% or 45.8%, respectively, relative to 50.6 days during 1986-2005 (Fig. 1a).
The annual mean intensity of high temperature during 1961-2005 shows an increasing trend of 0.07°C/10a. Similar to the frequency, the intensity will increase continuously until the 2050s under both pathways, RCP2.6 and RCP4.5. After the 2050s, the intensity will not increase under RCP2.6, but will still increase under RCP4.5. The intensity in the reference period was approximately equal to 1.6°C. Compared with the reference period, the intensity of high temperature is projected to increase by 1.2°C and 1.9°C at a global warming of 1.5°C and 2.0°C, respectively (Fig. 1b).
Changes in total mortality. As changing exposure and improved adaptation capacity change the risks of climate extremes, an adequate assessment of climate change impacts should take future socioeconomic development into account. Therefore, the population by age and gender, and the Gross Domestic Product (GDP) of 27 metropolises in China for the 21st century are projected under the framework of the Shared Socioeconomic Pathways (SSPs), which represent different climate strategies for mitigation and adaptation (Supplementary Fig. 4 and Supplementary Table 4). The SSPs describe a set of plausible alternative futures of societal development, which consider the effects of climate change and new climate policies. The SSPs include a pathway of a sustainable world (SSP1), a pathway of continuing historical trend (SSP2), a strongly fragmented world (SSP3), a highly unequal world (SSP4), and a growth-oriented world (SSP5) 41,42 . All five SSPs combined with RCP2.6 and RCP4.5 can produce ten plausible climatic-socioeconomic scenarios for the assessment of risks from high temperature. Additionally, GDP per capita in metropolises can be used as an indicator to evaluate the adaptability of different cities to high temperature ( Supplementary  Fig. 5).
On average, heat-related mortality in China metropolises was 32.1 per million by ensemble mean of the multiple GCMs in 1986-2005 (Fig. 2). Under the assumption that the socioeconomy remains stable at the 1986-2005 status, increasing frequency and intensity of high temperature will double the heatrelated mortality to 64.3 per million at global warming of 1.5°C, and even stronger increase to 85.5 per million at 2.0°C global warming (Supplementary Table 5).
However, exposure and vulnerability to high temperature are dynamic, and human adaptability to adverse climate is expected to increase with the socioeconomic development. When improved adaptation is integrated into assessment, interaction between the severity of high temperature and an increase in vulnerable population in the future will lead to increases in heatrelated mortality to 48.8-67.1 per million for 1.5°C global warming, across plausible development pathways, and to 59.2-81.3 per million for 2.0°C global warming (Fig. 2). That is to say, curbing the increase in global temperature to 1.5°C can reduce heat-related mortality in China metropolises by about 18% compared with 2.0°C.
Ignorance of contribution of adaptation actions could lead to substantial overestimation of climate change impacts. Without improved adaptation, heat-related mortality will be enlarged to 103.7-129.9 per million for 1.5°C global warming under various SSPs. Further increase in mortality to 137.3-169.9 per million was projected for 2.0°C warming (Fig. 2). For the urban population of 831 million in China, the extra heat-related mortality between 1.5°C and 2.0°C global warming will be in the range of 27.9-33.2 thousands, annually.  Table 6). Overall, female mortality was and will be continuously higher than for male, but the gap between genders is projected to be narrowed, due to the assumed changes in sex ratio in China from 105:100 in 1986-2005, for various SSPs, to (96-101):100 in 2060-2099.
If no improvement in adaptation capacity is assumed, mortality in the female and male population will be 71.  Fig. 6a).
For 1986-2005, heat-related mortality in the working age population was 7.0 per million and that of the non-working age population was 25.1 per million. With 1.5°C global warming, mortality in the working age population is projected to decrease significantly by 42.9%-60.0% to 2.8-4.1 per million. In contrast, mortality in the non-working age population is projected to increase significantly to 44.7-64.4 per million. This is an increase by 78.1%-156.6% compared to the reference period. With 2.0°C global warming, the mortality in the working age population will significantly decrease by 35.7%-57.1% to 3.0-4.5 per million. As for the non-working age population, it will significantly increase by 117.5%-211.6% to 54.7-78.2 per million. The increase of heatrelated mortality for the non-working age population and decrease for the working age population in China metropolises with the warming are mainly due to the projected demographic structure changes ( Fig. 3b and Supplementary Table 6). Under scenario without improved adaptation capacity, mortality will be 162.5%-167.9% higher for the working age population, and 87.1%-108.5% higher for the nonworking age population than projections with improved adaptability, at 1.5°C global warming. Mortality will be 224.4%-240.0% higher for the working age population and 100.6%-124.7% higher for the non-working age population, with the additional increase in global warming by 0.5°C (Supplementary Fig. 6b).

Discussion
With global warming, temperature extremes are likely to be more frequent, more intense, and longer lasting. In addition, demographics and adaptation capacities will change dramatically in future. The assessment of future changes in heatrelated mortality requires projections of the climate conditions, the population growth, the socioeconomic development, and consideration of improved adaptation. As far as we are aware, this is the first attempt to use locally defined concepts to investigate the relationship between high temperature and mortality for a large fraction of major cities in China. In this study, recorded cases from 27 metropolises are applied to deduce the threshold temperature for heat-related mortality. Furthermore, daily maximum temperature from 31 GCM outputs are combined with projected population under five SSPs to estimate mortality at 1.5°C and 2.0°C global warming,  Heat-related mortality increases above a certain threshold temperature with a non-linear relationship. This threshold temperature is the most critical information in preventing the health impacts of high temperature, as it is an indicator for initiating public health responses 5,43,44 . The threshold temperature is the temperature at which adverse health effects from heat begin to occur. The impacts are diverse for various categories, e.g. gender and age groups or geography 45 . Kan et al. investigated the relationship between daily mean temperature and mortality in Shanghai from January 2000 to December 2001 by using a generalized additive model, and found a gently sloping V-like relationship with the lowest mortality risk temperature of 26.7°C 46 . Another heat-related mortality study by Knowlton et al. found that the threshold temperature in New York is~23.1°C 13 . In our study, the gender-and age-specific mortality-inducing threshold temperature in Shanghai ranges around 29.7-31.4°C. In Beijing, which is located almost at the same latitude as New York, the threshold temperature is about 25.9-27.6°C.
Direct comparisons of the impact estimations are biased as different climate models, scenarios, downscaling methods, time periods, and population growth scenarios are used. To allow a rough comparison between this study and previous studies, we computed the changes in future heat-related mortality per million for scenarios not including improved adaptation capacity ( Fig. 2 and Supplementary Fig. 6). Our findings of increases in future heat-related mortality are broadly consistent with these assessments. We deduced an annual heat-related mortality of 32.1 per million in the reference period. No improvement in adaptation capacity is considered, the range of heat-related mortality will be 103.7-129.9 per million at 1.5°C global warming, and 137.3-169.9 per million at 2.0°C global warming. The mortality in China metropolises projected in our study is higher than in the United States for the last forty years of the 21 st century, which indicates a lower adaptation capacity in China than in the United States. Of course, other factors, such as the uncertainties in climate models, emission scenarios as well as baseline mortality rates, are also contributing to the differences in mortality estimations.
By incorporating future assumptions for an improved adaptability into assessment, a much lesser increase of mortality will be projected. Under improved adaptation capacity, annual heatrelated mortality is projected to be 48.8-67.1 per million at 1.5°C global warming, and 59.2-81.3 per million at 2.0°C global warming. That is to say, improved adaptation capacity will lead to 48.3-52.9% less mortality at 1.5°C, and 52.1-56.9% less mortality at 2.0°C global warming. Comparing with 2.0°C global warming, some 18% of mortality can be reduced in China metropolises by curbing temperature to 1.5°C.
It is a common assumption that heat-related mortality is more marked in the elderly and the female population, who are more vulnerable to the impact of high temperature than the adult and male population 36,48 . Some studies highlighted that females are at higher risks of dying or being sick during high temperature episodes 45,49 . According to the relative risk of specific temperature estimated by a distributed lag non-linear model, it is found that the threshold temperature for males is~0.8°C higher than for females, and for the working age population, it is 1.5°C higher than for the non-working age population (Supplementary  Table 3). With the warming, China will face adverse impacts due to the aging population. Our findings also suggest that heatrelated female mortality is much higher than for males at both global warming levels, but the gap between the mortality rates in males and females will slightly narrow in future, due to changes in the sex ratio in China.
The split of the working and non-working age population is projected to change quite seriously from 75.9%:24.1% in 1986-2005 to 43.8%:56.2% in 2060-2099. As the population structure will be extremely altered, the age-specific heat-related mortality will be different at 1.5°C global warming than at 2.0°C global warming. At 1.5°C global warming, the mortality in the working age population will be reduced by 42.9-60.0% relative to the reference period. On the contrary, the mortality in the nonworking age population will increase significantly by 78.1-156.6%. At 2.0°C global warming, the mortality in the working age population will be slightly higher than for 1.5°C global warming, while for the non-working age population mortality will be much higher with 2.0°C compared to 1.5°C.

Methods
Study area. In total, 27 major cities of China, i.e. metropolises, which include four municipalities (Beijing, Tianjin, Shanghai, and Chongqing) and most of the provincial capitals, are selected to project heat-related mortality under future climatic and socioeconomic scenarios. The population in each metropolis is above 2.0 million, and exceeds 10.0 million in Beijing, Chengdu, Chongqing, Guangzhou, Harbin, Shanghai, Shijiazhuang, and Tianjin. The total population and GDP of the 27 major cities were about 247.6 million people and 13.0 trillion CNY in 2010, which account for 18.6 and 29.7% of the national total, respectively (Supplementary Fig. 1 and Supplementary Table 1).
Mortality records. The daily mortality data in China metropolises during 2007-2013 were collected from the Chinese National Center for Chronic and Noncommunicable Disease Control and Prevention. The underlying cause of death was coded based on the 10th Revision of the International Statistical Classification of Diseases and Related Health Problems (ICD-10). Amongst, daily non-accidental mortality (ICD-10: A00-R99), mortality due to cardiovascular disease (I00-I99), respiratory disease (J00-J99), and so on were further categorized into four groups by age and gender: working age (age: 15-64 years) and non-working age (age: ≤14 and ≥65 years); female and male. Details of the mortality data can be found in a previous study by Yang et al. 27 .  Table 2). The GCM outputs were bias-corrected and downscaled statically to a regular geographical grid of 0.5°resolutions, based on observations, to show the GCMs have a good consistency in simulating high temperature in the major cities of China (Supplementary Fig. 2).  50,51 . The GDP in China under SSPs is projected with regionalized parameters using the Cobb-Douglas production model 52,53 , and is standardized to 2010 price level to maintain the homogeneity of data series. All the GDP and population are projected at the provincial scale first. Then, based on the county-level distribution of population and GDP in 2010, the area ratio method is applied to downscale population and GDP into the 0.5°resolution. Finally, the population and GDP within the boundaries of the city are summed. Distributed lag non-linear model. The temperature-mortality relationship is set up using a distributed lag non-linear model, which can describe complex nonlinear and lagged dependencies through the combination of the conventional exposure-response association and the additional lag-response association 54 .
A natural cubic B-spline of time with 8 degrees of freedom per year is applied to control long-term trends and to indicate the days of a week 55 . The lag-response association represents the temporal change in risk after a specific exposure, and estimates the distribution and delayed effects that cumulate across the lag period. We modeled the exposure-response curve with a quadratic B-spline with three internal knots placed at the 10th, 75th, and 90th percentiles of location-specific temperature distributions, and the lag-response curve with a natural cubic B-spline with an intercept and three internal knots placed at equally spaced values in the log scale. We extended the lag period to 10 days to include the long delay of the high temperature effects as it usually lasts around a week 36,[56][57][58] . The fitted metaanalytical model is used to derive the best linear unbiased prediction of the overall cumulative temperature and mortality association, and the minimum mortality temperature. We define the minimum mortality temperature as the threshold temperature.
Where E(Y t ) is the observed daily mortality at calendar day t; l refers to the maximum lag days, and Temp, l is the cross-basis matrix for the two dimensions of maximum temperature and lags; the natural cubic spline function NS() captures the non-linear relationship between the covariate (time) and mortality; Dow and Holiday are the dummy variables for the day of the week and public holiday; RR I,s is the relative risk corresponding to high temperature with certain intensity for metropolises, and greater or equal to 1; I is the intensity of high temperature, deduced by difference between daily maximum temperature and the minimum mortality temperature; and s represents the different metropolises. All analyses were performed using the R software Version 3.5.1 (R Foundation for Statistical Computing, Vienna, Austria) by using DLNM and MVMETA packages.
Projection of future heat-related mortality. Heat-related mortality at 1.5°C and 2.0°C global warming are projected by combining the simulated daily maximum temperature and the temperature-mortality relationship. We computed cityspecific heat-related mortality as follows: where s represents the different metropolises, I is the intensity of high temperature; M s is the daily heat-related mortality; Y s represents daily mortality rate per million in the observational period; POP s is the population; ERC I,s is the increase in relative risks along with intensification of high temperature, which is related to the improved adaptation capacity AC I (Supplementary Fig. 5).

Data availability
The dataset generated and analyzed during this study are available (with some institutional limitations) from the corresponding authors upon reasonable request. The source data underlying Figs. 1a-b, 2, and 3a-b are provided as a Source Data file.