Evaluating Japan’s revised heat-health warning system in the face of recent escalating heat stress

In 2021, the Japanese government changed the exposure indicator of the national heat-health warning system (HHWS) from air temperature (T air) to Wet Bulb Globe Temperature (T WBG), reflecting the growing concerns about the escalating humid heat stress. However, a clear validation of the advantages of using heat stress indicators (HSIs) that combine T air and humidity in heat alerts and heat-health applications are still being explored. Here, by using the latest epidemiological data (2015–2019) before the COVID-19 pandemic, we examined the rationality of the revised HHWS for 47 prefectures in Japan. Specifically, we investigated the predictive power of different HSIs in modeling mortality and morbidity caused by different diseases (e.g. all causes, circulatory, respiratory, and heatstroke) and age groups. Our findings revealed substantial differences among the HSIs in identifying periods of intense heat stress, potentially leading to differences in the activation dates of the HHWS if various indicators were employed. While HSIs exhibited comparable performances in modeling daily mortality, our analysis demonstrates distinct advantages in using T WBG for daily morbidity predictions, and the quasi-Akaike Information Criterion of T WBG is much lower than the previously used T air. The merits of T WBG are consistent in modeling all causes, non-external, as well as heatstroke-related morbidity. Overall, this study underscores the practicality of incorporating HSIs in heat stress early warnings and provides critical insights for refining the HHWS to mitigate health impacts from heat stress under future climate change.


Introduction
Climate change has greatly increased the frequency, duration, and intensity of heat waves in recent years, which has become a major concern for society (Feng et al 2021, Domeisen et al 2022, Gu et al 2022, Wu et al 2022, Masselot et al 2023, Yuan et al 2023).Compared to the global average temperature increase rate of 0.73 • C per century, the observed warming rate of Japan is much higher and reaches 1.28 • C per century (JMA 2021).Besides, for cities where populations gather, further temperature increases have been recorded due to urban heat island effects (JMA 2021).For instance, in the summer of 2022, Tokyo endured a record-breaking heatwave, with temperatures exceeding 35 • C for nine consecutive days.This event marked the most severe heatwave recorded in Tokyo since official temperature observations commenced in the 1870s (Witze 2022).In 2023, Japan broke temperature records with 106 out of 915 observation stations reporting the highest temperatures ever recorded.This resulted in July 2023 being documented as the hottest July in the country's history (Miyano 2023).These strong heat stress events resulted in serious heat-related illnesses, and in 2023 (May-September) alone, 91 467 people were transported by ambulance due to heat stroke across Japan, a substantial increase from the corresponding number of 71 029 people in 2022 (FDMA 2023b).
One of the effective measures to mitigate the health impacts of strong heat stress is the development of the heat-health warning system (HHWS).HHWS aims to accurately forecast heat stress and encourage citizens to take appropriate actions (e.g.limit outdoor activities, keep hydrated, and check vulnerable individuals) to lower heat-related health risks.To address the escalating challenge of heat stress, HHWS has been adopted in various cities across North America and Europe, although the specific features of these systems may vary to accommodate local climate conditions and the vulnerability of the population (Kotharkar and Ghosh 2022).In Japan, a similar HHWS was established in 2011, where a heat warning is triggered when the maximum temperature reaches or exceeds 35 • C (Oka et al 2023a).
Due to the sweat and evaporative cooling system of the human body, the human-perceived heat stress is not only controlled by air temperature (T air ), but also largely influenced by humidity, wind speed, and incident radiation (de Freitas and Grigorieva 2017, Sherwood 2018).Numerous studies have been investigating human tolerance to extremely humid and hot conditions in recent years, contributing to a deeper understanding of human-perceived heat stress (Chakraborty et al 2022, Guo et al 2022, Lu and Romps 2023, Lu and Romps 2023, Sherwood and Huber 2010, Vecellio et al 2022, 2023, Zhang et al 2023).These studies propose the use of heat stress indicators (HSIs), which comprehensively consider multiple climate variables, with a particular emphasis on the combined impact of T air and humidity.As a result of this research progress, the Japanese government has made notable revisions to the HHWS in Japan (Oka et al 2023a).Specifically, the exposure indicator has shifted from T air to Wet Bulb Globe Temperature (T WBG ), and the updated system has been carried out for the whole country since 2021 (Liljegren et al 2008, JMA & MOE 2020).T WBG considers the joint effect of T air , humidity, wind speed, and solar radiation, and is also widely used in the United States to indicate heat category for outdoor physical activities (Liljegren et al 2008).Under this updated HHWS in Japan, a heat warning is issued when the daily maximum T WBG is predicted to reach or exceed 33 • C (Oka et al 2023a).In 2023, a total of 1232 heat-related alerts were issued across the country (MOE 2023b).Besides, different T WBG ranges are also used to provide suggestions for daily life and outdoor activities.For example, when T WBG reaches the level of 28 • C-31 • C, the HHWS suggests avoiding exposure to sunlight when going out and being careful of the rise of the indoor temperature.For outdoor activities, intense exercise that can easily raise body temperature should be avoided, like endurance running, and taking breaks every 10-20 min to replenish water and salt (MOE 2023a).
The relationship between elevated ambient temperatures and human mortality and morbidity has been extensively explored in research, encompassing various aspects such as heatwave intensity and duration (Guo et al 2017), temperature fluctuations (Wu et al 2022, Wen et al 2023), and the impact of urban heat islands (Huang et al 2023), among others.However, the merits of using HSIs that combine T air and humidity in heat alerts have been debatable and a clear validation of the advantages of HSIs in heathealth applications is still being explored (Baldwin et al 2023).In addition, the health burden of heat extremes on city residents is not only limited to heatstroke, but also various heat-vulnerable diseases and the corresponding mortality (Vicedo-Cabrera et al 2021).To date, there has been limited exploration of the comparative efficacy of T WBG in place of T air and other widely used HSIs in predicting broader health outcomes, such as all-cause mortality and morbidity in Japan.
Here, by using the latest epidemiological data for the 47 prefectures in Japan as well as nine widely used HSIs calculated with high temporal resolutions, the predictive power of these HSIs for heat-related mortality and morbidity are well investigated, and the merits of T WBG are examined.Two main objectives of this study include: Firstly, we aim to unveil the differences in the information provided by these HSIs regarding 'Hot Days' .We specifically examine and compare the timing of the annual hottest ten days as detected by different HSIs for the period 2015-2019.Secondly, we assess the predictive performance of various HSIs in predicting mortality and morbidity in Japan, considering factors such as specific causes, age groups, and geographical variations.To our knowledge, this is the most comprehensive study to date on the association between humid heat stress and health outcomes in Japan, which provides an important reference for the future improvement of national HHWS.

Data
This study integrates both health and weather data.The health data encompass daily mortality and morbidity (Emergency Ambulance Dispatch, EAD) spanning 2015-2019 across Japan's 47 prefectures.Daily mortality data was obtained from the Ministry of Health, Labor, and Welfare of Japan, and the daily morbidity data was obtained from the Fire and Disaster Management Agency (FDMA) of the Ministry of Internal Affairs and Communication of Japan.These datasets include comprehensive records of all-cause, circulatory (ICD (International Classification of Diseases)-10: I00-I99), respiratory (ICD-10: J00-J99), and non-external (ICD-10: A00-R99) related mortality and morbidity, categorized by all-age and elderly individuals (age ⩾ 65).Past studies have indicated that circulatory and respiratory diseases are notably susceptible to intense heat stress, therefore these two types of diseases are studied here particularly (Ebi et al 2021).Additionally, the daily EAD data for heatstroke incidents (ICD-10: T67.0-3, T67.5-8) (Oka et al 2023a), obtained from the FDMA for the same 47 prefectures during the period 2015-2019, is incorporated into the analysis (FDMA 2023a).
The ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts-(ECMWF) (Hersbach et al 2023) is used to calculate the HSIs.Specifically, the hourly 2 m air temperature, 2 m dewpoint temperature, 10 m wind speed, surface pressure, and surface downward solar radiation, are utilized, spanning the period from 2015 to 2019.The grid (∼31 km × 31 km) centered on each prefecture's capital city is used to represent the climate conditions for the prefecture.Previous research has validated the reliability of the ERA5 datasets as suitable surrogates for station data in epidemiological studies (Mistry et al 2022).Given the focus on heat stress, only observations during the warm season (June-September) are analyzed in this research.Descriptive statistics detailing population demographics, average mortality, and morbidity rates, as well as climate conditions across the 47 prefectures in Japan, are provided in table S1 in the supplementary.(Brode et al 2012), and Heat Index (HI) (Rothfusz 1990) are calculated at an hourly resolution based on ERA5 data.The T WBG is calculated utilizing the R (version 4.0.3)package 'HeatStress' , and the UTCI is calculated utilizing the Python (version 3.7.12)package 'pythermalcomfort' .The other HSIs are calculated based on the formulas provided in the corresponding references.Then, the daily mean value of these HSIs is obtained by averaging the hourly values for a day, considering the local time zone of Japan.Previous studies have indicated that heat-related fatalities are influenced not only by daytime heat exposure but also significantly by factors such as elevated nighttime temperatures (He et al 2022, Kim et al 2023).Consequently, in this study, we utilize the daily mean values of HSIs, which are proposed to offer a more comprehensive representation of overall daily heat exposure and exhibit greater predictive capability (Guo et al 2011).The detailed information and the inputs of each HSI are provided in table S2.

Exposure-response analysis
The distributed lag non-linear models (DLNMs) (Gasparrini et al 2010) are used to model the exposure-response association between heat stress exposure and health outcomes.DLNMs (R package 'dlnm') are well-established methods that are capable of modeling complex non-linear and lagged associations often found in heat-health studies.We established the association between the daily mean value of these HSIs (as well as T air ) with the health outcomes using a quasi-Poisson regression: where the outcome Y t corresponds to daily counts of mortality or morbidity assumed to follow a Poisson distribution with overdispersion; the function f(x t ;θ) represents the bi-dimensional exposure-lag-response association, which is modeled through a combination of two functions defined within a cross-basis term; s(t;β) represents the temporal components which encompass confounders changing slowly over time (i.e.seasonal and long-term trends); and h p (z pt ;γ p ) specifies the contributions of other p numbers of confounders varying on a daily basis.The determination of model settings was drawn from relevant studies (Vicedo-Cabrera et al 2021, Oka et al 2023b).Specifically, we model the exposureresponse curve using a natural cubic spline function with two internal knots positioned at the 50th and 90th percentiles (with equal spacing when fitting for heatstroke) of the warm season indicator distribution.For the lag-response curve, we employ a natural spline function with two internal knots equally spaced values on a 10 day lag (a 7 day lag when fitting for heatstroke) on the log scale.To account for seasonality, we utilize a natural spline function with 4 degrees of freedom for the day of the year, while a natural spline function of time with one knot is employed to capture long-term trends in mortality/morbidity.Additionally, the model incorporates indicators to account for intra-week variation and holidays in mortality/morbidity.
The impact of heat stress exposure on mortality and morbidity is assessed using the relative risk (RR, Gasparrini et al 2010).This measure quantifies the probability of an outcome, such as morbidity, occurring in a group exposed to a specific environmental factor in comparison to the probability in a group not exposed to that factor.In this study, we established the reference T air /HSI value representing the point of minimum mortality/morbidity risk, denoted where RR = 1.Subsequently, we computed the incremental RR for varying T air /HSI concerning the risk relative to this reference value.A higher RR value indicates an Q Guo et al increased risk of health-related outcomes associated with the exposure.

Competitive performances of HSIs
The quasi-Akaike Information Criterion (qAIC) (Gasparrini et al 2010) serves as a statistical tool employed to assess the predictive performance of each exposure indicator in relation to health outcomes.This criterion allows for the comparison of the goodness of fit among different statistical models, with lower qAIC values indicating superior model fit.
To ensure comparability, all indicators (i.e.HSIs and T air ) are fitted using identical DLNM structures and based on equivalent data samples for each prefecture (Armstrong et al 2019).
The competitive performance of exposure indicators is defined as: where the 'best-fit indicator' signifies the exposure indicator with the lowest qAIC value.In essence, when the difference between an indicator's qAIC and the best-fit indicator is less than 2 (Burnham andAnderson 2002, Burnham et al 2010), we consider this indicator to perform well, signifying competitiveness and proximity to the best-performing model.

Overlap of the annual hottest ten days detected by different HSIs
As HSIs are devised by various researchers based on different assumptions, these indicators often encompass distinct ranges and exhibit varying sensitivity to changes in T air and Relative Humidity (RH) (de Freitas and Grigorieva 2017, Sherwood 2018).Among the nine HSIs utilized, indicators like T w , T s , and T WBG exhibit heightened sensitivity to RH fluctuations (figures 1(a)-(c)).Conversely, HI, UTCI, and APT showcase lower sensitivity to RH and are more influenced by T air (figures 1(g)-(i)).Consequently, due to the joint impact of T air and RH variations, significant disparities arise regarding days with extreme heat stress when assessed using different HSIs (figure 2).For instance, our analysis indicates notable deviations in the annual hottest ten days identified by T WBG in comparison to those identified by T air , particularly noticeable in southern Japan, where the overlap rates stand at only 60%-70% for certain prefectures (figure 2(a)).This underscores how the choice of exposure indicators, such as T air and T WBG , significantly influences the HHWS in determining when to issue a heat alert.Among all HSIs, Tw exhibits the lowest overlap rate with T air in identifying hot days, whereas HI demonstrates a relatively higher overlap (figure 2(b)).Furthermore, the consistent trend across all HSIs analyzed is a lower overlap rate observed in southern Japan (i.e.prefectures with ID 24-47 in figure 2(b)), attributed to the region's warmer and more humid climate conditions.

Competitive performances of HSIs in predicting different health outcomes
We investigated the response of mortality and morbidity to the change in heat stress exposure based on the HSIs and T air .Tokyo and Osaka, two megacities located in northern and southern Japan, respectively, were examined as examples (figure 3).Our findings reveal that in both cities, once the exposure indicator surpasses a specific threshold, there is a noticeable increase in RR, with morbidity exhibiting a relatively higher rate of increase (figures 3(c) and (d)).Additionally, variations exist in the exposure-response curves among HSIs and T air , indicating differences in the predictive abilities of heat stress.
We compared the competitive performances of HSIs and T air in modeling mortality (figure S1) and morbidity (figure 4) across 47 prefectures in Japan, considering all causes and all age groups.Generally, different HSIs and T air demonstrated similar results in modeling all-cause mortality, with HI, Hx, and T sWBG showing relatively better performances compared to other exposure indicators (figures S1(c), (g), (h)).The qAIC differed little among these HSIs and T air , mostly within 10 (figure S2).However, significant disparities were observed in competitive performances among different indicators for morbidity (figure 4).Indicators like ESI, Hx, T sWBG , T s , and T w displayed comparatively poor performances and lacked competitive results across all prefectures (figures 4(f)-(j)).Conversely, T air , T WBG , and UTCI exhibited better performances, with T WBG demonstrating competitive performances across a larger number of prefectures (figures 4(a), (b) and (d)).Moreover, the qAIC differences in modeling morbidity (figure S3) were notably larger than those observed for mortality (figure S2), sometimes reaching up to 150.These substantial qAIC differences highlight significant disparities in the predictive power of morbidity among HSIs and T air .
We summarized the performance of these exposure indicators in determining their competitive performances in simulating mortality (figure S4) and morbidity (figure 5) by causes and age groups across the 47 prefectures.Interestingly, all indicators displayed similar performances in simulating daily mortality under different classifications (figure S4) and daily morbidity caused by circulatory and respiratory diseases (figures 5(b), (c), (e) and (f)).However, in terms of predicting daily all-cause morbidity, there was a considerable disparity in the performance of different indicators, with T WBG generally outperforming others, alongside T air and UTCI, which also

The strength of T WBG in predicting daily morbidity
To understand the substantial performance variations among different HSIs in predicting all-cause morbidity (figures 5(a) and (d)) in contrast to other health outcomes (figures S4 and 5(b), (c), (e), (f)), we checked the seasonality of these health outcomes (figures S5 and S6 for Tokyo and Osaka as examples).Unlike other health outcomes that are typically higher in winter and lower in summer (figures S5(a)-(c), (e), (f) and S6(a)-(c), (e), (f)), all-cause morbidity cases peak up in both winter (i.e.January and December) and summer (i.e.July and August) (figures S5(d) and S6(d)).This suggests that all-cause morbidity may be more sensitive to heat stress during summer compared to other health outcomes.Consequently, T WBG is likely to offer enhanced predictive capabilities for these heat-sensitive health outcomes.
To further assess T WBG 's comparative effectiveness in modeling morbidity across distinct categories, we opted to analyze two specific health outcomes: heatstroke morbidity, which is directly attributed to external heat stress, and non-external morbidity (involving acute illnesses excluding external causes such as accident and heatstroke), which albeit is not immediate consequences but also greatly associated with intense heat.Our objective was to evaluate the performances of all HSIs and T air regarding these two direct and indirect health outcomes influenced by heat stress, which are well-represented and can provide comprehensive insights.Seasonality of heatstroke and non-external morbidity show clear increases during summer, indicating their higher susceptibility to heat stress (figure S7).The exposureresponse association also demonstrates a substantial RR increase for both health outcomes under heat stress, especially for heatstroke morbidity (figure S8).
Comparing the predictive performances of all indicators for heatstroke and non-external morbidity  S1, 1-47 corresponds to the direction from north to south) in Japan.
(figures S9 and S10), T WBG emerges as largely advantageous for both health outcomes (figures S9(b) and S10(b)).Furthermore, the qAIC disparities among various indicators are substantial, reaching up to 400 for both health outcomes (figures S11 and S12).These findings highlight considerable variation in the predictive abilities of different HSIs, with T WBG consistently demonstrating superior performance in predicting heatstroke and non-external morbidity.
Subsequently, we summarized the competitive performances of each indicator in forecasting heatstroke and non-external morbidity across all 47 prefectures, categorized by age groups (figures 6(a)-(f)).Across these analyses, T WBG emerged as the topperforming indicator for predicting heatstroke and non-external morbidity in Japan.Notably, T WBG displayed competitive performance in over half of the prefectures, supported by substantial qAIC differences compared to T air (figures 6(g) and (h)), further advocating for the adoption of T WBG in the updated HHWS as the preferred exposure indicator over the former T air .

Discussion
In comparing the response between morbidity and mortality data in Japan, our analysis revealed a lower sensitivity of mortality to heat stress, indicated by a smaller RR (figures 3(a) and (b)).This diminished susceptibility of population mortality to heat stress has been noticeable in recent decades within Japan.For instance, Chung et al (2018) reported that the minimum mortality temperature (MMT) in Japan increased from 23.2 • C to 28.7 • C from 1972 to 2012, and the heat-related mortality risk also showed a clear decrease.This trend could be explained by population adaptation to heat stress.Besides, socioeconomic factors, such as increased air conditioning usage, could also influence human mortality susceptibility to heat.Sera et al (2020) observed a clear reduction in heat-related excess deaths in several countries, including Canada, Japan, and Spain, attributed to increased air conditioning prevalence.These factors potentially contribute to the limited differences in qAIC across various HSIs and T air (figure S2), leading to overall similar performances in modeling daily mortality (figure S4).
Our study utilized ERA5 reanalysis data, merging various sources like radiosonde, satellite-based, and in-situ observations, to establish associations with epidemiological observations.Research  in-situ data obtained from the Japan Meteorological Agency and revealed good agreements (figure S13), which affirms the reliability of ERA5 data.Moreover, considering the mountainous terrain of Japan and the concentration of the population in capital cities of prefectures (figure S14, CIESIN 2018), using the ERA5 grids centered around these capital cities could provide a good representation of the climate exposure for the prefectural population.
To define competitive performances among HSIs, we employed a ∆qAIC < 2 criterion.As a sensitivity analysis, we also presented results based on ∆qAIC < 7 to assess the robustness of our findings (Burnham andAnderson 2002, Burnham et al 2010).These sensitivity tests consistently indicated similar relative performances among HSIs, affirming T WBG as the superior indicator in modeling allcause, non-external, and heatstroke-related morbidity (figures S15-S17), which confirmed the robustness of our findings.
Our findings demonstrate the validity of adopting T WBG as the exposure indicator for Japan's revised HHWS, which is more informative in indicating the health impacts due to heat stress among other metrics.These results contribute valuable insights to the ongoing discussion surrounding the influence of various meteorological variables, particularly humidity, on heat-related health outcomes (Baldwin et al 2023).However, given the diverse climatic landscapes and varying levels of adaptive capacity among populations globally, we exercise caution in extrapolating our findings to other regions.Moreover, for many other HSIs tested in this study, the inclusion of multiple climate variables alongside T air does not yield improved performance in modeling health outcomes, possibly due to unrealistic assumptions, simplified calculations, or improper representation of climate variables.Therefore, T air remains a simple and reliable indicator, particularly in cases where climate data are limited, or the benefits of using HSIs are not validated for specific locations.To facilitate the establishment of effective HHWS and tailored adaptation strategies for different locales, we recommend conducting similar population-scale exposure-response analyses tailored Q Guo et al   to the unique climatic contexts of each region, thereby identifying the most effective HSIs for localized implementation.

Conclusions
Due to the rising concern of the severe humid heat stress under climate change, the Japanese government has changed the exposure indicator of the national HHWS from air temperature (T air ) to Wet Bulb Globe Temperature (T WBG ) in 2021.In this study, by using the latest epidemiological data in Japan as well as high-resolution ERA5 reanalysis data, we comprehensively examined the performance of T WBG in predicting daily mortality and morbidity under different disease and age classifications.The performances of T WBG are evaluated against commonly used T air and several widely used HSIs.Our findings highlight significant disparities among different HSIs in identifying heat stress events (referred to as 'hot days'), in southern Japan.Hence, the careful selection of indicators for the HHWS holds substantial practical importance.HSIs generally demonstrate similar performances in modeling daily mortality.when modeling all-cause, non-external, and heat stroke morbidity, notable disparities in the indicators' performances emerge.Among all HSIs and T air , T WBG exhibits distinct advantages in modeling these morbidity measures, with considerably lower qAIC values compared to T air .This study confirms that employing HSIs, with careful selection, surpasses the commonly used T air in designing HHWS and enhances the system's predictive capabilities.Overall, our results underscore the crucial role of indicator selection in HHWS design and improvement to alleviate the health impacts of severe heat stress under future climate change.
support the findings of this study are available upon reasonable request from the authors.

Figure 1 .
Figure 1.The sensitivity of nine heat stress indicators (HSIs) to changes in air temperature (T air , • C) and relative humidity (RH, %).(a)-(i), the shaded color represents the value of the corresponding HSI, and the contour represents the reference value of Tw, which is most sensitive to change in RH.The nine HSIs are listed in descending order of the sensitivity to RH (from a to i: Tw (most sensitive to RH) to HI (least sensitive to RH)).

Figure 2 .
Figure 2. The overlap rate (%) of the timing of the hottest ten days of the warm season (Jun-Sep, averaged for 2015-2019) detected by different HSIs and T air .(a), the rate in the timing of the hottest ten days of warm season detected by TWBG and T air for 47 prefectures in Japan.(b), the overlap rate in the timing of the hottest ten days of the warm season detected by each HSI and T air for 47 prefectures (denoted as ID 1-47 in tableS1, 1-47 corresponds to the direction from north to south) in Japan.
by Mistry et al (2022) demonstrated the comparable accuracy of ERA5 and ERA5-Land data with in-situ climate observations for epidemiological studies.Additionally, compared with station data, which is usually limited in data length, locations, and types of climate variables, the reanalysis data can provide sufficient climate variables with consecutive time series and spatial coverage.Nowadays, reanalysis data like ERA5 have been extensively employed in diverse epidemiological analyses (Urban et al 2021, Huang et al 2023, Lo et al 2023), as a powerful surrogate of station data.To ensure the reliability of our analysis, we validated ERA5's T air and RH against corresponding

Figure 3 .
Figure 3. Exposure-response association for Tokyo and Osaka obtained by each HIS and T air .(a), the relative risk (RR) of mortality associated with exposure indicators (i.e.HSIs and T air ) in Tokyo.(b), the RR of mortality associated with exposure indicators in Osaka.(c), The RR of morbidity with exposure indicators in Tokyo.(d), the RR of morbidity associated with exposure indicators in Osaka.Shading represents the 95% empirical confidence interval.The mortality and morbidity here are all causes and all ages.The values of HSIs are shown in quantile values (%).

Figure 4 .
Figure 4.The spatial distribution of the competitive performances (orange shaded) in modeling morbidity for T air (a) and each HSI (b)-(j).The results are based on all causes and all ages morbidity.

Figure 5 .
Figure 5.The number of prefectures (maximum 47) that T air or HSIs (x-axis) show competitive performance in modeling morbidity for all 47 prefectures in Japan.(a)-(f), results for different categories of morbidity.The dashed horizontal line represents the median value across 47 prefectures.

Figure 6 .
Figure 6.The performances of T air and HSIs in modeling heatstroke and non-external morbidity.(a)-(c), The numbers of prefectures (maximum 47) that T air or HSIs (x-axis) show competitive performance in modeling heatstroke morbidity.(d)-(f), The numbers of prefectures (maximum 47) that T air or HSIs (x-axis) show competitive performance in modeling non-external morbidity.The dashed horizontal line in (a)-(f) represents the median value across 47 prefectures.(g)-(h), The qAIC difference between T air and TWBG (the former minus the latter) in modeling heatstroke and non-external morbidity, respectively.Young and elderly represent individuals with age <65 years and ⩾65 years, respectively.