Interventional applications of a Stroke Heat Risk Prediction Model produce health benefits

Abstract

Although heat exposure increases stroke risk, targeted individualized interventions remain limited. This study develops and validates a Stroke Heat Risk Grading Prediction Model for precision intervention using 28,116 stroke deaths from 304 Chinese counties. Meteorological and stroke mortality data from 2013–2018 are analyzed with time-series methods, revealing a nonlinear temperature-mortality relationship. Four risk levels are established and validated using 2019–2022 data through case-crossover and time-series analyses considering sex, age, and geography. At the highest risk level of our model, stroke mortality increases by 13.8% in the general population and 16.4% in older adults, whereas the China Meteorological Administration warning system poorly predicts stroke mortality. Interventions guided by our model achieve nearly a two-fold increase in the proportion of avoidable heat-attributable excess deaths compared to existing approaches. These findings support this model as a digital tool to mitigate heat-related stroke risk under climate change.

Although heat is a major stroke risk factor, targeted intervention tools are lacking, so we developed and evaluated a Stroke Heat Risk Model for personalized prevention. Here, the authors show that it yields three times the preventable heat-attributable stroke deaths of current heat warning systems.

Introduction

According to the 2019 Global Burden of Disease (GBD) report, stroke ranks as the second leading cause of global disease burden and the primary cause of disease burden in China¹, which accounts for over a quarter of the global stroke population, with approximately 28.76 million affected individuals¹. Global stroke prevention guidelines published in The Lancet indicate that heat exposure² significantly exacerbates stroke risk, as stroke is particularly sensitive to heat. The 2021 GBD report further emphasizes that heat was responsible for approximately 2 million disability-adjusted life years (DALYs) from stroke globally³. Despite this, effective interventions addressing heat as a risk factor remain lacking. While traditional metabolic and behavioral risk factors have been standardized in stroke prevention protocols, stroke incidence and mortality rates continue to rise. As such, managing heat as a stroke risk factor, a currently overlooked aspect in prevention, has become increasingly critical.

At present, heatwave forecasting and warning systems are the most commonly employed methods for alerting heat risks^4,5 globally. However, these systems are limited in their ability to predict the health impacts of heat, particularly regarding specific diseases such as stroke. They typically provide alerts based on temperature alone, without accounting for the heat health effects. Moreover, a uniform threshold applied nationwide fails to consider the variability in heat-related health effects due to factors such as age, sex, and regional climate, thereby reducing the effectiveness of heatwave warnings in preventing heat-sensitive diseases like stroke. Currently, no tools exist for precise heat health management tailored to stroke prevention and intervention.

In this work, we developed a Stroke Heat Risk Grading Prediction Model using meteorological and stroke mortality data from 304 counties in China, spanning from 2013 to 2022. By comparing this model with the existing heatwave warning grading system employed by the China Meteorological Administration (CMA), the study clarifies the practical value of the Stroke Heat Risk Grading Prediction Model for health management and stroke prevention.

Results

Summary statistics for stroke mortality

Supplementary Fig. 1 illustrates the characteristics and environmental factors associated with stroke mortality, based on data collected from 304 counties across China between 2013 and 2022. Table 1 shows that the dataset comprised 15,263 samples from 2013 to 2018 (training set) and 12,853 samples from 2019 to 2022 (validation set), totaling 28,116 samples. Among the stroke-related deaths, ischemic stroke accounted for 63.6%, significantly surpassing the 36.4% attributed to hemorrhagic stroke. Notably, the proportion of ischemic stroke increased from 62.2% to 65.4%, while hemorrhagic stroke decreased from 37.8% to 34.6% (P = 6.14 × 10⁻⁶).

Table 1.

Descriptive analysis of characteristics and environmental factors associated with stroke mortality for the 304 counties in China

Characteristics	All (n = 28116)	2013–2018 (n = 15263) a	2019–2022 (n = 12853) a	P-value
Motality, n(%)
Ischemic stroke	17894 (63.6)	9487 (62.2)	8407 (65.4)	6.14 × 10−6
Hemorrhagic stroke	10222 (36.4)	5776 (37.8)	4446 (34.6)
Region, n(%)
South	15865 (56.4)	8116 (53.2)	7749 (60.3)	3.28 × 10−7
North	12251 (43.6)	7147 (46.8)	5104 (39.7)
Sex, n(%)
Male	14168 (50.4)	8270 (54.2)	5898 (45.9)	1.83 × 10−7
Female	13948 (49.6)	6993 (45.8)	6955 (54.1)
Age, n(%)
< 65	2316 (8.2)	1879 (12.3)	437 (3.4)	1.76 × 10−7
≥ 65	25800 (91.8)	13384 (87.7)	12416 (96.6)
Environmental factors, median (p25, p75)
Temperature (°C)	25.70 (23.42, 28.16)	25.95 (23.42, 28.16)	25.52 (23.18, 27.6)	1.49 × 10−6
Wind speed (m/s)	1.98 (1.38, 2.80)	1.98 (1.39, 2.80)	1.97 (1.38, 2.79)	0.205
Relative humidity (%)	73.10 (59.94, 82.38)	78.48 (69.49, 84.89)	78.51 (68.84, 85.1)	0.707
PM2.5 (μg/m3)	35.52 (21.28, 59.39)	33 (22, 48.5)	18.16 (10.99, 30.2)	2.51 × 10−5
O3 (μg/m3)	82.88 (61.41, 110.69)	83.5 (57, 121)	123.57 (95.34, 155.74)	7.03 × 10−6

^aMeteorological and stroke mortality data from 304 counties in China were divided into a training set covering 2013 to 2018 and an evaluation set covering 2019 to 2022. Continuous environmental factors are presented as median with 25th and 75th percentiles (p25, p75). Differences between the 2013–2018 and 2019–2022 datasets were assessed using Pearson’s chi-squared tests for categorical variables, and Mann-Whitney U tests for continuous variables, with Bonferroni correction applied for multiple comparisons. All tests were two-sided.

Table 1 indicates that individuals aged 65 years or older comprised 91.8% of all stroke-related deaths, with only 8.2% occurring among those under 65 years. This age disparity became more pronounced in the 2019–2022 period, with the proportion of stroke mortality among those aged ≥ 65 years rising significantly from 87.7% to 96.6% (P = 1.76 × 10^–7).

Establishment of the stroke heat risk grading prediction model

Figure 1a shows that high-temperature exposure was associated with an increasing trend in stroke mortality risk among individuals younger than 65 years; however, the effect was not statistically significant for either sex. In contrast, for individuals aged 65 years or older, high-temperature exposure significantly increased the stroke mortality risk. Under extreme high-temperature conditions (95th percentile), the stroke mortality risk was slightly higher for females (relative risks [RR]: 1.437, 95% confidence intervals [CI]: 1.142, 1.807) compared to males (RR: 1.426, 95% CI: 1.163, 1.750).

Fig. 1. Establishment and validation of the Stroke Heat Risk Grading Prediction Model.

a The exposure-response relationships between temperature and stroke mortality, derived from time-series analysis (n = 15,263 independent stroke mortality cases). b The effect estimates for stroke mortality under various risk levels among the general population and the elderly population in one proposed strategy (Low: 0% ≤ excess risk [ER] < 5%; Moderate: 5% ≤ ER < 10%; High: 10% ≤ ER < 40%; Extremely high: ER ≥ 40%) for the Stroke Heat Risk Grading Prediction Model, derived from case-crossover design and conditional logistic regression (n = 12,853 independent stroke mortality cases). c Grading principles for heat-related health risks in the Stroke Heat Risk Grading Prediction Model for cities in different climate zones. The solid line in (a) represents the predicted relative risk (RR) of stroke mortality, and the shaded band represents the 95% confidence interval (CI). Points in (b) represent the estimated odds ratios (ORs), and vertical lines represent the 95% confidence interval (CI). Source data are provided as a Source Data file.

Supplementary Table 1 presents three distinct risk grading strategies were then employed. The exposure-response relationship between relative temperature and stroke mortality risk revealed a maximum RR value of approximately 1.5. Initially, the curve displayed a linear trend; however, the rate of increase in stroke mortality risk accelerated as relative temperature rose. Based on comprehensive and reasonable considerations and the identification of inflection points corresponding to the most pronounced changes in risk, we determined a relative risk of 1.4 (corresponding to an excess risk [ER] of 40%) as the threshold for categorizing the “extremely high-risk” category and developed three distinct risk grading strategies.

Validation of the Stroke Heat Risk Grading Prediction Model

Figure 1b and Supplementary Figs. 2–13 present the validation of different strategies for the Stroke Heat Risk Grading Prediction Model based on case-crossover analysis. As shown in Fig. 1b, by applying the extremely high-risk threshold (ER ≥ 40%) and implementing incremental decreases with 30% and 5% thresholds (Low: 0% ≤ ER < 5%; Moderate: 5% ≤ ER < 10%; High: 10% ≤ ER < 40%; Extremely high: ER ≥ 40%), an increase in stroke mortality risk was observed with higher risk levels. These findings informed the determination of the optimal grading model. Furthermore, other grading strategies did not demonstrate comparable sensitivity of risk grading, as shown in the Supplementary Figs. 2–13.

Supplementary Figs. 2–13 depict the effect estimates for stroke, ischemic stroke and hemorrhagic stroke mortality across various risk levels among the general population and the elderly population in different proposed strategies for various lag days. In the optimal grading model, the lag days for stroke, ischemic stroke, and hemorrhagic stroke were determined to be lag2, lag2, and lag1, respectively, based on the highest odds ratios (ORs) observed at the extremely high-risk level in the general population. Specifically, the ORs for moderate-, high-, and extremely high-intensity risk levels, along with their corresponding mortality risks, were 0.985 (95% CI: 0.934, 1.039), 1.047 (95% CI: 1.005, 1.091), and 1.138 (95% CI: 1.079, 1.200) for stroke, 0.969 (95% CI: 0.905, 1.038), 1.010 (95% CI: 0.958, 1.065), and 1.193 (95% CI: 1.117, 1.275) for ischemic stroke, and 1.019 (95% CI: 0.935, 1.110), 1.068 (95% CI: 0.998, 1.142), and 1.097 (95% CI: 1.002, 1.202) for hemorrhagic stroke in the general population. In the elderly population, the optimal grading model also demonstrated superior sensitivity in risk grading. In addition, Supplementary Fig. 14 further demonstrated that stroke mortality risk increased progressively with higher risk levels in the optimal grading model, based on time-series analysis.

Figure 1c details the grading framework for heat-related health risks within the Stroke Heat Risk Grading Prediction Model. Risk was classified into four levels based on ER values and corresponding warning colors: low risk (ER: 0%–5%, green), moderate risk (ER: 5%–10%, blue), high risk (ER: 10%–40%, orange), and extremely high risk (ER ≥ 40%, red). These categories visually and quantitatively illustrated the escalating impact of heat exposure on stroke mortality, with higher ER values and warmer colors indicating greater risk.

Furthermore, Fig. 1c demonstrates the application of the Stroke Heat Risk Grading Prediction Model across various cities in different climatic zones, including Guangzhou (tropical), Hefei (subtropical), and Beijing (temperate). The figure shows how the model’s heat-related health risk grading applies to different age and sex groups within each city.

Evaluation of the Stroke Heat Risk Grading Prediction Model

Figure 2a depicts the ORs for various risk levels within the Stroke Heat Risk Grading Prediction Model. Figure 2b further demonstrates that even in cases classified as “no risk” under the CMA heatwave warning grading model, but identified as “at risk” by the Stroke Heat Risk Grading Prediction Model, there was a significant increase in stroke mortality risk on the same day (lag0).

Fig. 2. The effect estimates for stroke mortality under various risk levels among the general population and the elderly population.

a Odds ratios (ORs) estimated for different risk levels in the Stroke Heat Risk Grading Prediction Model intervention scenario, derived from case-crossover design and conditional logistic regression (n = 12,853 independent stroke mortality cases) and (b) under the China Meteorological Administration (CMA) heatwave warning grading model intervention scenario, derived from case-crossover design and conditional logistic regression (n = 12,853 independent stroke mortality cases). Points in (a, b) represent the estimated odds ratios (ORs), and vertical lines represent the 95% confidence interval (CI). Source data are provided as a Source Data file.

Table 2 indicates that the estimated attributable number of stroke deaths associated with heat exposure was 1775 for stroke, 1144 for ischemic stroke, and 578 for hemorrhagic stroke. Under the Stroke Heat Risk Grading Prediction Model intervention scenario, the estimated excess deaths that could be avoided were estimated at 871, 575, and 301, respectively. In contrast, the CMA heatwave warning grading model intervention scenario estimated significantly fewer avoided excess deaths: 304 for stroke, 193 for ischemic stroke, and 110 for hemorrhagic stroke.

Table 2.

Estimated heat-attributed deaths, and the number and proportion of excess deaths attributable to heat that could be avoided under the Stroke Heat Risk Grading Prediction Model intervention scenario and the CMA heatwave warning grading model intervention scenario from 2019 to 2022 in China

Disease	Estimated Heat-attributed deaths	Stroke Heat Risk Grading Prediction Model		CMA Heatwave Warning Grading Model
		Estimated population	Estimated percentage	Estimated population	Estimated percentage
Stroke
General population	1775	871	49.1%	304	17.1%
Elderly population	1370	828	60.4%	249	18.2%
Ischemic stroke
General population	1144	575	50.3%	193	16.9%
Elderly population	920	562	61.1%	170	18.5%
Hemorrhagic stroke
General population	578	301	52.1%	110	19%
Elderly population	482	280	58.1%	81	16.8%

Figure 3a demonstrates that the estimated proportion of avoided excess deaths attributable to heat under the Stroke Heat Risk Grading Prediction Model intervention scenario was notably higher, reaching 49% for stroke, 50% for ischemic stroke, and 52% for hemorrhagic stroke. In contrast, the proportions under the CMA existing heatwave grading model intervention scenario were significantly lower, at 17%, 17%, and 19%, respectively.

Fig. 3. Evaluation of the Stroke Heat Risk Grading Prediction Model’s applicability.

a The estimated proportion of avoided excess deaths attributable to heat under the Stroke Heat Risk Grading Prediction Model intervention scenario and the China Meteorological Administration (CMA) existing heatwave grading model intervention scenario in the general population (n = 12,853 independent stroke mortality cases), and (b) the elderly population (n = 12,416 independent stroke mortality cases).

Figure 3b highlights that, in the elderly population, the estimated proportion of avoided excess deaths attributable to heat under the Stroke Heat Risk Grading Prediction Model intervention scenario increased, with stroke rising to 60%, ischemic stroke to 61%, and hemorrhagic stroke to 58%. Conversely, the proportions under the CMA heatwave warning grading model intervention scenario exhibited minimal change.

Figure 4 illustrates the demo of a stroke heat-health risk early warning system based on the Stroke Heat Risk Grading Prediction Model, tailored to local climatic characteristics. The system could disseminate real-time, graded heat-health risk levels to the public via mobile applications and We-chat official accounts. By integrating meteorological forecast data, it would facilitate several-day prospective estimations of stroke heat risks, categorized into four distinct risk levels, thus providing the high-risk groups with accessible and actionable information.

Fig. 4. Mobile app interface of the Stroke Heat Risk Grading Prediction Model.

The model provides real-time heat-health risk warnings at four levels and prospective risk estimations based on meteorological forecasts and individual characteristics (location, sex, and age).

Discussion

This study analyzed over 20,000 meteorological and mortality records from 304 counties in China between 2013 and 2022, leading to the development and evaluation of a Stroke Heat Risk Grading Prediction Model. The findings indicate that the constructed model effectively identifies health risks. Compared to the lowest risk level, the highest risk level of this model predicted a 13.8% increase in stroke mortality risk for the general population and a 16.4% increase for the elderly population. Under the intervention scenario using the Stroke Heat Risk Grading Prediction Model, the health benefits were nearly twice greater than those achieved by the existing heatwave warning grading model used by the CMA. This research introduces an innovative approach to constructing heat risk prediction models, which is vital for stroke prevention and management.

Heat-related stroke mortality exhibits age-specific heterogeneity, highlighting the necessity for age-differentiated risk prediction. Several multi-center studies conducted in China on the general population have found that temperatures exceeding the 95th percentile (defined as extreme heat) significantly increase the RR of stroke mortality by ~ 10%–15% compared to the MMP^6,7. Our findings are consistent with these studies, demonstrating a clear trend of increased stroke mortality risk due to heat exposure. Notably, this effect is more pronounced in the elderly population, where stroke risk rises by ~ 30%. The heightened vulnerability in older individuals is attributed to impaired autophagy and heat shock protein response (HSR), as well as exacerbated inflammatory and coagulation responses under heat stress, all of which contribute to elevated stroke risk^8–10. As a result, this study specifically differentiates between younger and older individuals, underscores the importance of age-related differences, and develops a stroke risk stratification prediction model.

The innovative customized heat-health risk grading prediction model specifically targeting stroke has been developed, offering a more effective approach to managing heat-related stroke risk compared to the existing heatwave warning grading model issued by the CMA. In terms of the rationality of risk level settings guided by health risks, the model categorizes heat-health risks into four levels: low, medium, high, and extremely high. Stroke mortality risk exhibits a progressive increase across these levels, with approximately a 10% rise per level, demonstrating the relevance and significance of the risk level framework. In contrast, the current heatwave warning grading model, which includes three levels (yellow, orange, and red), lacks sufficient differentiation in predicting health risks. Although the first two levels indicate significant health risks, they do not differ substantively, and the third level fails to predict health risks effectively. Regarding the highest risk level’s capacity for health identification, the model predicts a stroke mortality risk of up to 13.8% in the general population under extremely high risk and as high as 16.4% among the elderly. In comparison, the highest level of the current heatwave warning grading model does not effectively predict stroke mortality risk. Furthermore, regarding the management of heat risk factors, previous studies have established that when temperatures exceed the Minimum Mortality Temperature (MMT), heat becomes a significant health risk^11–13. Our model encompasses all temperature ranges above the MMT, whereas the current heatwave warning grading model focuses solely on high temperatures, neglecting the health risks associated with medium-high temperatures. In these overlooked ranges, stroke mortality risk remains as high as 5.6%, or up to 10% among the elderly.

The Stroke Heat Risk Grading Prediction Model demonstrates substantially greater health benefits. When implemented alongside the CMA’s current heatwave warning grading model and accompanied by appropriate protective measures, our model can prevent 49 excess stroke deaths per 100 individuals due to heat exposure, compared to just 17 excess deaths per 100 individuals with the existing model. This results in a nearly twofold increase in health benefits. Notably, among the elderly population, intervention scenarios based on our heat-health risk grading prediction model could prevent up to 60 excess stroke deaths per 100 elderly individuals, while the current model shows negligible improvements. These findings underscore the significant advantages of integrating risk prediction with health grading, particularly for the high-risk elderly population, offering more accurate assessments and effectively mitigating heat-related mortality risk.

In the future, the stroke heat-health risk grading prediction model can be seamlessly integrated with mobile phones, wearable devices (e.g., smartwatches), and remote medical equipment (such as electronic blood pressure monitors and electrocardiogram [ECG] devices). This integration would allow for the automatic identification of geographic location via GPS and the input of individual information such as age and sex, enabling the provision of graded forecasts of stroke heat-health risks for the current day and the subsequent seven days. In addition, the model will generate timely risk warnings and offer personalized health intervention recommendations, including advice on diet, outdoor activities, home temperature management, and medical consultations. These capabilities will empower users to take preventive actions in advance, reducing the incidence and mortality risks of heat-induced stroke.

Conducting a modeling exercise for the Stroke Heat Risk Grading Prediction Model provides significant health benefits and represents a critical step prior to implementation. Modeling is a standard practice in many health-related studies, particularly when evaluating innovative interventions or prediction tools before implementation. This approach is scientifically sound and widely accepted in the medical research community. For example, a study demonstrated that the PREVENT cardiovascular disease risk assessment equation from the American Heart Association exhibited excellent discriminatory ability after simulated validation using a representative population across the United States¹⁴. Another study used model-based simulations and subsequent validations to confirm that an all-cause mortality prediction model based on social determinants of health outperformed traditional assessment methods in predictive performance¹⁵. The model-building process that incorporates simulation-based evaluations ensures that the model is thoroughly tested before being applied, providing robust scientific support for policy-making. This strategy is widely regarded as a key tool in health policy development due to its ability to inform and justify evidence-based decisions.

The stroke heat risk grading prediction model can serve as a pioneering digital health intervention tool for predicting and warning against heat-related health risks. By utilizing GPS for location tracking and requiring only basic personal information such as age and sex, the tool forecasts stroke-related heat-health risks for the current day and the next seven days. It also provides tailored recommendations for dietary adjustments, outdoor activity limitations, home temperature control, and medical consultations. This tool is particularly valuable in primary, secondary, and tertiary prevention strategies. In primary prevention, it serves the general population, highlighting stroke—a heat-sensitive condition—as a critical indicator of heat-related health impacts. The model will educate the public on the risks associated with heat exposure and promote awareness, helping individuals mitigate the negative effects of extreme heat. In secondary prevention, the model targets high-risk groups, complementing existing stroke risk screening tools such as FAST (Face-Arm-Speech-Time), RACE (Rapid Arterial Occlusion Evaluation scale), G-FAST (Gaze-Face-Arm-Speech-Time), and CG-FAST (Conveniently-Grasped Field Assessment Stroke Triage)^16–19. This integration facilitates tailored interventions and management for individuals identified as high-risk through screening, effectively preventing stroke incidents. In tertiary prevention, the model focuses on patients with stroke, specifically addressing the risk of recurrence. By incorporating the heat-health risk grading model into existing digital health-based rehabilitation and treatment programs, heat-related risk factors can be managed, ultimately reducing the likelihood of stroke recurrence.

Several limitations of this study warrant consideration. Firstly, the analysis of mortality data from 304 counties across China revealed limited representation of western regions, primarily due to smaller population sizes and insufficient monitoring infrastructure. However, the coverage is relatively comprehensive when compared to previous nationwide multicenter studies. To ensure the reliability of the data, only counties with a mortality rate exceeding 5‰ were included. Secondly, the study focused primarily on mortality as the health endpoint for evaluating heat impacts, as mortality data typically offers superior quality and is more widely available. This approach provided a robust and objective measure for model validation, ensuring the credibility of the findings in assessing heat-related health effects. Nonetheless, other significant health outcomes, such as morbidity, hospital admissions, and outpatient visits, were not included in the analysis. Future research should incorporate a broader range of health endpoints to enable more nuanced evaluations and more effective public health strategies. Thirdly, due to the limited spatial and temporal coverage of local meteorological station data, the study utilized reanalyzed meteorological data from the ECMWF. While direct monitoring data would be preferable for its localized accuracy, the ECMWF dataset, which is widely recognized for its comprehensive coverage and extensive validation in previous studies, provided a reliable alternative. Fourth, due to data availability constraints, we employed two distinct analytical methods across different time periods: time-series analysis for 2013–2018 and case-crossover analysis for 2019–2022. During the earlier period (2013–2018), only daily population-level mortality data were accessible, which required the use of time-series methods to establish exposure–response relationships and develop different grading strategies. To enable a more individualized assessment, we specifically collected person-level data for 2019–2022, which allowed us to apply a case-crossover design for validation and assess the applicability of the Stroke Heat Risk Grading Prediction Model. Previous studies have demonstrated that time-series and case-crossover analyses produce comparable results in evaluating heat-related health risks^20,21. Our strategy thus represents a deliberate methodological progression from population-level to individual-level analysis, directly addressing the limitations of aggregated data. Validation results indicate that our model performs well in predicting individualized heat-related health risks, supporting its applicability in personalized early warning systems. This approach effectively mitigates data limitations while preserving scientific rigor.

In conclusion, this study presents the development of a four-level stroke heat-health risk prediction model, categorizing risks into low, medium, high, and extremely high levels. The model integrates variables such as age, region, and sex, offering an innovative approach to managing heat-related stroke risk, particularly in the context of climate change. Compared to the current heatwave warning grading system employed by the CMA, this model demonstrates enhanced predictive capabilities and significantly greater health benefits. In addition, it can be transformed into a digital health intervention tool, facilitating early heat risk grading and predictions. This tool is suitable for primary, secondary, and tertiary prevention of heat-related stroke risk, particularly benefiting high-risk groups such as the elderly.

Methods

This study complies with all relevant ethical regulations and was reviewed and approved by the Institutional Ethics Committee of the National Institute of Environmental Health, Chinese Center for Disease Control and Prevention (Approval No. NIEH202410). The committee waived the requirement for written informed consent because the study used existing, de-identified public health surveillance data.

Study design

A comprehensive modeling approach for the Stroke Heat Risk Grading Prediction Model has been developed, encompassing the construction of graded prediction models, model validation, and the evaluation of intervention under different scenarios. Firstly, meteorological and stroke mortality data from 304 counties in China were partitioned into a training set (2013–2018) and an evaluation set (2019–2022). Secondly, time-series analysis was conducted to identify the nonlinear relationship between temperature-induced heat effects and stroke mortality. Utilizing the principle of abrupt risk changes, various grading strategies were formulated to develop a series of heat risk grading prediction models. Thirdly, the validation set was used to identify the optimal prediction model through case-crossover and time-series analyses. Fourthly, scenario analysis was employed to assess the practical application value of the model. Specifically, the health-risk discriminative abilities and avoidable heat-related excess deaths of two scenarios were compared: one based on the Stroke Heat Risk Grading Prediction Model and the other on the current heatwave warning grading model used by the CMA (Fig. 5).

Fig. 5. Flowchart for the construction and evaluation of the Stroke Heat Risk Grading Prediction Model.

The framework includes model establishment using training data (2013–2018), validation using evaluation data (2019–2022), and evaluation of applicability by comparing with the current CMA heatwave warning system.

Data source

Heat health risk studies are typically conducted during high-temperature periods (i.e., the summer months)^22–25. This methodological choice ensures a more accurate capture of the distinctive health effects of high temperatures and is widely regarded as the standard practice for assessing heat-related health risks. Our study collected mortality data, meteorological data, and air pollution data during the summer months from 2013 to 2022.

Mortality data, classified according to the International Classification of Diseases, 10th Revision (ICD-10), were sourced from the China Cause-of-Death Surveillance dataset, maintained by the Chronic Non-Communicable Diseases Prevention and Control Center of the China Center for Disease Control and Prevention. To ensure data quality, counties with a mortality rate exceeding 5‰ were selected, resulting in the inclusion of 304 counties in the study. Stroke mortality case data (ICD-10: I60–I64) from June to August of each year between 2013 and 2022 were collected, with each case’s data encompassing age, gender, cause of death (ICD-10 code), and residential address.

Meteorological data were obtained from the hourly gridded Reanalysis v5-Land (ERA5-LAND) dataset provided by the European Center for Medium-Range Weather Forecasts (ECMWF). This dataset, accessible at https://cds.climate.copernicus.eu/datasets/reanalysis-era5-land?tab=overview, was used to extract data specific to the China region. The key variables included temperature at a 2 m height, relative humidity, and the horizontal and vertical components of wind speed. The dataset’s spatial grid resolution is 0·1° × 0·1°, corresponding to approximately 9 km. Wind speed was calculated as the square root of the sum of the squares of the horizontal and vertical wind speed components, as described in Formula 1.

$$Wind=\sqrt{{U\_Wind}^2+{V\_Wind}^2}$$ 1

In this formula, U_Wind and V_Wind represent the horizontal wind speed component (m/s) and the vertical wind speed component (m/s), respectively.

Daily aggregated variables were computed from the available hourly grid-scale data, including daily mean temperature (Tmean), daily mean relative humidity (RHmean), and daily mean wind speed (Wmean). The grid center points were matched to the respective counties using ArcGIS software version 10.2, which facilitated the aggregation of grid-scale data to the county level. For counties with multiple grid center points, the average value of all relevant grids was used as the representative value. In cases where a county did not contain a grid center point, data from the nearest grid center to the county’s centroid were assigned. The meteorological data from June to August of each year between 2013 and 2022 were collected.

PM_2.5 and O₃ concentration data were sourced from monitoring sites operated by the China National Environmental Monitoring Center. The original data consisted of hourly measurements collected from each monitoring site, with daily averages computed by aggregating the hourly values. If more than 25% (i.e., over 4 h) of hourly values were missing for a given day, the daily average for that site was considered missing. To estimate the average daily concentrations of PM_2.5 and O₃ for each county from June to August during 2013–2022, we calculated the mean of the daily concentrations from all sites within the county. In counties lacking dedicated monitoring sites, data from the nearest available monitoring site to the county’s center were utilized.

Establishment of the Stroke Heat Risk Grading Prediction Model

Sub-datasets from 2013 to 2018 were selected. A two-stage time series analysis was conducted to investigate the relationship between daily relative temperature and mortality risk. For the first stage, and daily mortality records were systematically categorized into four distinct groups: males under 65 years, females under 65 years, males aged 65 years and above, and females aged 65 years and above. Daily average temperature data from 2013 to 2018 were transformed into daily relative temperature, expressed as percentiles of historical temperature distributions. A quasi-Poisson regression model, combined with a Distributed Lag Nonlinear Model (DLNM), was applied to evaluate the association between relative temperature and stroke mortality across different counties and subgroups. The primary model accounted for several covariates, including daily mean relative humidity, PM_2.5 concentrations, O₃ concentrations, wind speed, temporal trends, and day-of-week effects, as shown in Formula 2.

$$\mathrm{Log}\left[E\left(u_t\right)\right]=\alpha +\mathrm{cb}\left({Temp}_t,lag\right)+\mathrm{ns}\left(Time,df\right)+\mathrm{ns}\left({RH}_t,df\right)+\mathrm{ns}\left({WS}_t,df\right)+{PM}_{2\cdot 5}+O_3+dow$$ 2

In this formula, u_t represents the number of stroke mortalities on day t, and α denotes the intercept. The term cb refers to the cross-basis function, while lag indicates the time lag. The natural cubic spline function ns is employed with a specified degree of freedom (df). The variable Temp_t denotes the daily average temperature for the current day and the previous t days, given that studies have shown the health impacts of heat are more pronounced in the short term, with a maximum lag time of 3 days⁶. Variables RH_t and WS_t represent the daily mean relative humidity and wind speed, respectively. $\mathrm{dow}$ represents the day of the week.

Temperature was modeled using a natural cubic spline with three internal knots, positioned at the 10th, 75th, and 90th percentiles of the county-specific temperature distribution. The lag structure was modeled using a natural cubic spline with 4 degrees of freedom, with two internal knots placed at equally spaced log-values of the lags from 0 to 3 days, allowing for greater flexibility at shorter delays. To account for long-term trends and seasonality, a natural cubic spline with 7 degrees of freedom per year was used for time, while a natural cubic spline with 3 degrees of freedom was applied for both relative humidity and wind speed. In addition, an indicator variable for the day of the week was included to capture complex temporal patterns.

In the second stage, the overall associations between temperature and stroke mortality were analyzed. A meta-analysis with a random-effects model was conducted to address the substantial heterogeneity across counties, pooling the specific estimates from the first stage. Separate meta-analyses were performed for subgroups based on age and gender. The RRs and 95%CIs of stroke mortality were computed by comparing the minimum mortality percentile (MMP) with the lowest stroke mortality risk.

Finally, the nonlinear exposure-response relationship curves between the high-temperature range of daily relative temperature (above the MMP) and stroke mortality risk were derived, considering a cumulative 3-day lag period across different age and gender subgroups.

Using a threshold selection procedure based on derivative analysis²⁶, we employed the first derivative of the exposure-response curve to identify inflection points corresponding to the most pronounced changes in risk, as shown in the Formula 3.

$$f\left(Temp\right)=\frac{df\left(Temp\right)}{dTemp}$$ 3

Temp represents the relative temperature, expressed as percentiles of the historical temperature distribution. f(Temp) denotes the cumulative exposure-response relationship between relative temperature and stroke mortality risk as estimated by the Formula 2, while f’(Temp) represents the first derivative of f(Temp), describing the rate at which mortality risk changes with increasing temperature.

Based on the results of derivative analysis and considering effect magnitude, consistency with internationally validated warning systems, exposure-response curve characteristics, and practical applicability, we initially determined the threshold for extremely high risk, as detailed in the Supplementary methods. Subsequent risk categories were then established by proportionally lowering this threshold across different decrement dimensions to identify the optimal stratification scheme. This strategy—initially determining the highest risk threshold and then defining subsequent levels through systematic decrements—was commonly used in risk stratification frameworks. A comparable approach was adopted by the World Health Organization (WHO) in developing global air quality standards, where the strictest Air Quality Guideline (AQG) level was initially set as the highest health protection standard, followed by the establishment of a tiered framework through successive decrements²⁷. Consequently, different risk grading strategies were then employed.

Validation of the Stroke Heat Risk Grading Prediction Model

External validation is essential to confirm the model’s effectiveness across diverse populations and settings. Current research methodologies generally require independent external validation to establish the model’s validity and scientific basis. For instance, the chronic disease polygenic risk score model underwent external validation to ensure its applicability across populations with different ancestral backgrounds, while the coronary artery disease multi-ethnic risk score model was independently validated in various populations, highlighting the critical role of external validation in supporting cross-ethnic applicability^28,29. In contrast, many warning model studies in the field of extreme weather events have focused primarily on model development and often lack essential validation, or have not conducted any validation at all^30,31.

For validation, data from 2019 to 2022 were selected as the evaluation set. We began by establishing a grading framework comprising “low risk”, “moderate risk”, “high risk” and “extremely high risk” categories, incorporating three distinct grading strategies. Using case-crossover and time-series analyses, we estimated the effects of various risk levels on mortality from stroke, ischemic stroke, and hemorrhagic stroke across different lag days (lag 0-3) for both the general and elderly populations. Case-crossover analysis, a study design used to assess the effects of transient exposures on acute events, compares each individual to themselves over different time periods. In this approach, the exposure status during the “hazard period”—immediately before an acute event such as a stroke—is compared to exposure during one or more “control periods,” when the event did not occur. In this study, control periods were selected from different time points for the same individual, such as the same day of the week in adjacent weeks, to control for temporal trends. By comparing within subjects, the case-crossover design inherently accounts for all time-invariant confounders, including genetic factors and long-term lifestyle habits.

The ORs corresponding to the graded risk levels (moderate, high, and extremely high) were compared to the low-risk level. We selected the lag day with the highest ORs at the extremely high-risk level, and assessed whether there are progressive transitions in mortality risk, specifically evaluating how effectively each grading strategy captures the increase in risk as warning levels escalate. Through this comprehensive validation, the final Stroke Heat Risk Grading Prediction Model was defined, ensuring an accurate representation of the relationship between varying risk levels and associated health outcomes.

Based on these findings, we developed corresponding health recommendations for different warning levels as appropriate interventions, as outlined in the Supplementary Table 2. We outlined a specific approach for developing recommendations.

Step 1, we developed a theoretical framework to systematically guide appropriate interventions. This framework comprises two fundamental dimensions: first, recommendations were stratified according to the Stroke Heat Risk Grading Prediction Model, categorizing risk into four levels—Low, Moderate, High, and Extremely High Risk. Second, tailored recommendations were developed for each of the four risk levels, addressing two key components: exposure-related and health-related factors. The recommendations increase in intensity with rising risk levels, providing proportionate and actionable guidance.

Step 2, recommendations to reduce heat exposure emphasized staying indoors during periods of extreme heat, maintaining a cool and well-ventilated indoor environment, minimizing abrupt indoor–outdoor temperature transitions, and practicing sun protection when outdoors. In addition, recommendations for health maintenance included ensuring adequate hydration, maintaining a balanced diet and sufficient rest, adhering to prescribed medications, and seeking timely medical attention when experiencing heat-related symptoms such as dizziness or fatigue.

Step 3, recommendation intensity increased with escalating risk levels. At lower levels, recommendations emphasized general preparedness, maintaining a stable indoor environment, and sustaining healthy daily routines. With increasing risk levels, recommendations expanded to include the avoidance of outdoor activities, enhancement of indoor cooling, assurance of adequate hydration and rest, and increased monitoring for heat-related symptoms. At the highest risk level, additional emphasis was placed on strengthening home preparedness, enhanced symptom awareness, and timely medical attention.

These recommendations were informed by extensive reviews of authoritative resources as shown in the Supplementary Table 3, notably the WHO’s “Heatwaves: Risks and Responses”, “Heatwaves and Health: Guidance on Warning-System Development” and “Public Health Advice on Preventing Health Effects of Heat”, as well as the “Guidelines for Public Health Protection Against Heatwaves” and the “Health Recommendations and Graded Early Warning Approach for Heat Health Risk” issued by China’s National Disease Control and Prevention Administration^32–36. It is important to highlight that the latter two guidelines were specifically developed by our research team. To further ensure clinical validity and applicability, we referenced the “2024 AHA/ASA Guideline for the Primary Prevention of Stroke” and the “2022 Chinese Guideline for the Secondary Prevention of Ischemic Stroke and Transient Ischemic Attack”^37,38. The recommendations were finalized after expert consultations and interdisciplinary deliberations involving clinical specialists.

Evaluation of the Stroke Heat Risk Grading Prediction Model’s applicability

At present, the only officially implemented early warning system for meteorological factors is the conventional CMA system, which primarily focuses on atmospheric phenomena and presents notable limitations in addressing individual-level health protection. In contrast, our individualized model adopted a human-centered approach by shifting the focus from environmental conditions to personal vulnerability. Specifically targeting populations at high risk of stroke, the model facilitated forecasting and early warning management of heat-related health risks. Currently, there is no dedicated tool for this purpose, and high-risk individuals can only rely on general weather forecasts or the meteorological warnings of the CMA warning system. To address this gap, we compared our individualized model with the CMA warning system by designing two intervention scenarios, aiming to evaluate the applicability of the Stroke Heat Risk Grading Prediction Model and to provide empirical support for its broader implementation.

Two intervention scenarios were developed: one based on the Stroke Heat Risk Grading Prediction Model and the other on the existing heatwave warning grading system used by the CMA. In both scenarios, it was assumed that the Stroke Heat Risk Grading Prediction Model and the CMA’s heatwave warning system would be applied concurrently at the population level, with corresponding protective measures implemented. To assess the health-risk discriminative abilities and avoidable heat-related excess deaths of the Stroke Heat Risk Grading Prediction Model, stroke mortality risks across different risk levels in both models were compared, focusing particularly on the excess stroke deaths attributable to heat that could potentially be mitigated.

To evaluate the discriminative ability of the Stroke Heat Risk Grading Prediction Model in distinguishing varying degrees of health risks across different warning levels and assess whether health risks increase with escalating warning levels, we conducted the following analyses using data from 2019 to 2022. With the lowest risk level from the Stroke Heat Risk Grading Prediction Model serving as the reference, we estimated the ORs associated with the condition when the CMA did not issue a heat risk warning while the Stroke Heat Risk Grading Prediction Model did. In addition, we estimated the impact of varying warning levels from both the CMA’s heatwave warning system and the Stroke Heat Risk Grading Prediction Model on stroke mortality among the general and elderly populations.

The number and proportion of excess deaths attributable to heat that could be prevented under two intervention scenarios could be utilized by a retrospective approach^39–42. This approach utilized nonlinear exposure-response relationship curves that correlate high-temperature ranges of daily relative temperature with stroke mortality risk, as established through case-crossover analysis. Furthermore, estimations were made regarding the excess number of stroke deaths (stroke, ischemic stroke, and hemorrhagic stroke) that could be prevented under the two intervention scenarios, as quantified by Formula 4.

$${\sum PAN}_{j,k}=N_{j,k}\times \frac{\left({OR}_{j,k}-1\right)}{{OR}_{j,k}}$$ 4

In this formula, PAN_j,k represents the estimated excess deaths attributable to heat that could be avoided for disease k related to the risk level j, N_j,k denotes the total daily number of deaths for disease k during the days corresponding to risk level j, and OR_j,k refers to the estimated mortality risk of disease k associated with risk level j.

To further assess the health benefits of the Stroke Heat Risk Grading Prediction Model, the proportion of excess deaths attributable to heat that could be avoided under the two intervention scenarios was calculated. This proportion was derived by dividing the total estimated avoided excess deaths in both scenarios by the estimated number of stroke deaths attributable to heat.

Statistical analyses were performed using the “survival” and “dlnm” packages in R (version 3.6.3). The results were presented as ORs with 95%CI, reflecting the mortality risk associated with different graded risk levels in the Stroke Heat Risk Grading Model.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

T.L. designed the study, guided analysis, revised and reviewed the manuscript. J.Z. performed data cleaning, data analysis, and manuscript writing. M.Z. and Q.D. conducted data curation. Q.S., R.M., C.Z., and K.L. reviewed and edited the manuscript. All authors contributed to the manuscript. All authors have given approval to the final version of the manuscript.

Peer review

Peer review information

Nature Communications thanks Ewout Steyerberg and Jun Yang for their contribution to the peer review of this work. A peer review file is available.

Data availability

The data generated in this study are available under restricted access for regulations imposed by the ethics committee, and the raw datasets used in this study are not publicly accessible. However, researchers can contact the corresponding author (litiantian@nieh.chinacdc.cn) to request access to the data for reproducibility analysis or collaborative research. Requests will be answered within 12 weeks. The exposure data for air pollution used in this study was available from the National Urban Air Quality Real-Time Release Platform (https://air.cnemc.cn:18007/); the meteorological exposure data used in this study were sourced from the ERA5-Land reanalysis dataset released by European Center for Medium-Range Weather Forecasts (https://cds.climate.copernicus.eu/datasets/reanalysis-era5-land?tab=overview). Source data are provided in this paper.

Code availability

Code used in this study is available online at (https://github.com/Zhangjingwei-cdc-tj/Stroke_heat_risk_grading_prediction_model_NatureComm), under the MIT license. The specific version of the code associated with this publication is archived in Zenodo and is accessible via 10.5281/zenodo.17718307⁴³.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-68815-4.