Critical Temperature Thresholds for Identifying Vulnerability to Heat‐Related Excess Cardiovascular Morbidity and Mortality

Abstract

Background

Extreme heat is a well‐established environmental hazard linked to elevated cardiovascular disease (CVD) morbidity and mortality, yet few studies evaluate temperature effects across multiple temporal scales or identify community‐specific vulnerability.

Methods

We analyzed neighborhood‐level CVD deaths and emergency department visits linked to sociodemographic characteristics and high‐resolution temperature and humidity estimates. Generalized additive models with smooth splines for temperature, humidity, age, and time estimated excess heat‐related rates across temporal scales. Principal component analysis and k‐means clustering classified Chicago community areas by multidimensional heat vulnerability.

Results

Higher temperatures were significantly associated with increased CVD and coronary heart disease mortality across warm‐season, monthly, and daily scales but were not associated with cardiovascular emergency department visits. Peak warm‐season thresholds for all CVD mortality occurred at 25.6 °C, corresponding to 20.7 excess deaths per 100 000 (SD 20.3; P<0.001). Daily peaks occurred at 39.5 °C with 0.048 excess deaths per 100 000 per day (SD 0.068; P<0.001), and a 0 to 3‐day lag peak at 38 °C produced 0.049 excess deaths per 100 000 (SD 0.133; P<0.001). Coronary heart disease mortality showed similar patterns, with warm‐season peaks at 27.8 °C (9.19 per 100 000; P=0.004). No statistically significant associations were observed for myocardial infarction or stroke mortality. Principal component analysis and clustering identified 3 vulnerability profiles driven by socioeconomic disadvantage, racial and ethnic composition, heat exposure, and humidity.

Conclusions

Temperature thresholds for cardiovascular mortality vary across temporal scales and CVD subtypes, with strongest associations for all CVD and coronary heart disease mortality. Integrating temperature–mortality relationships with community vulnerability profiles may support targeted heat warning systems and neighborhood‐specific adaptation strategies.

Nonstandard Abbreviations and Acronyms

GAM

generalized additive model

PC1/PC2

principal component 1/principal component 2

PCA

principal component analysis

Tmax

maximum temperature

Tmean

mean temperature

Research Perspective.

What Is New?

• This study identifies clear, nonlinear temperature thresholds at which cardiovascular mortality (potentially driven by coronary heart disease) rises across warm‐season, monthly, and daily time scales, while showing no corresponding associations for emergency department visits, myocardial infarction, or stroke.

• By integrating 12 years of mortality and emergency department data with high‐resolution climate metrics and sociodemographics, we reveal 3 distinct vulnerability clusters driven by socioeconomic disadvantage, racial and ethnic composition, and environmental exposures that meaningfully stratify heat‐related mortality risk.

What Question Should Be Addressed Next?

• Future research should determine whether these temperature–mortality thresholds and vulnerability clusters generalize to other cities and climates, and whether targeted, community‐specific heat‐mitigation strategies can reduce cardiovascular deaths during extreme heat events.

Cardiovascular disease (CVD) is the leading cause of morbidity and mortality in the United States and heat exposure increases the burden of disease. After initial declines in CVD mortality over the past decade, the age‐adjusted mortality rate climbed 9.3% to nearly 500 per 100 000 by 2022. ¹ , ² CVD event incidence is affected by heat exposure, which is expected to increase in severity and duration because of climate change. ³ , ⁴ Excess cardiovascular deaths due to extreme heat are estimated to increase 161% to 232% by 2065 in the United States. ⁵ Urban areas with a large heat island effect are the most vulnerable to heat‐related CVD, with Chicago projected to have the largest burden of annual CVD morbidity and mortality of all major US cities by 2090. ⁶ , ⁷ , ⁸ Identifying the areas of Chicago that are most susceptible to heat‐related CVD will be important for resource allocation and preventive efforts in the face of consistently higher‐than‐average temperatures.

There are no standard definitions for extreme heat or heatwaves, and little work has been done to concomitantly assess both acute and long‐term impacts of heat on communities. ⁹ Many studies have characterized the impact of heat exposure on CVD hospitalizations and mortality. ¹⁰ , ¹¹ , ¹² However, these studies are often unable to identify vulnerable areas and microclimates because they are generally at a large geographic level, use low resolution temperature estimate methods, or fail to capture the effects of both short‐ and long‐term heat exposure on health outcomes. Prior studies have estimated the relative risk of CVD using predefined percentile thresholds for temperature that provide subjective associations that potentially miss larger patterns. ¹⁰ , ¹³ Likewise, the pathophysiological mechanisms of CVD in the presence of heat exposure manifest differently depending on CVD subtype, necessitating a spectrum of temperature thresholds to determine risk.

In this study, we evaluated the impact of temperature during warm months on excess CVD emergency department (ED) visits and mortality rates within Chicago community areas for various temporal scales. The purpose of this study was to isolate the effect of cumulative heat exposure on CVD morbidity and mortality using high resolution temperature estimates, identify population‐based temperature thresholds for CVD and its subtypes, and characterize community‐level vulnerability to short‐ and long‐term heat exposure.

METHODS

In order to minimize the possibility of unintentionally sharing information that can be used to reidentify private information, all code and a subset of the data generated for this study are available on GitHub and can be accessed at https://github.com/petergraffy/heat_cvd_chicago.

This was an institutional review board‐approved retrospective ecological time series study performed in collaboration with 7 academic medical centers in Chicago and the Illinois and Chicago Departments of Public Health. This study was approved by an institutional review committee under waiver of informed consent due to retrospective nature of deidentified health record data. We included decedents and ED patients age ≥18 years with an address that spatially corresponded to a Chicago community area between 2011 and 2022 for months May through September.

Chicago has 77 community areas that were originally established in 1920 and whose boundaries have remained unchanged since 1980. The most recent geographic shapefile for the community areas is publicly available through the City of Chicago Data Portal (https://data.cityofchicago.org/). Annual sociodemographic estimates of community area characteristics were queried from the ACS (American Community Survey) 5‐year tables. The variables selected include annual estimates of total population, education level, racial composition, sex distribution, employment status, median income, and median age for each community area across the entire 2011 to 2022 study period.

Daymet is a publicly available product developed by Oak Ridge National Laboratory that contains daily estimates of maximum temperature (Tmax), minimum temperature, and vapor pressure at a 1 km² spatial resolution across North America. Daily mean estimates of Tmax, minimum temperature, and vapor pressure were aggregated for each community area for the duration of the study period by spatially joining the community area shapefile with a raster brick of Daymet estimates and calculating the weighted mean of Daymet grid cells within each respective community area. Daily mean temperature and relative humidity values were calculated using a formula specific to Daymet estimates containing Tmax, minimum temperature, vapor pressure, and 3 constants, as described in prior literature. ¹⁴

To account for chronic environmental conditions that may confound heat–mortality relationships, we incorporated annual concentrations of fine particulate matter and nitrogen dioxide, as well as annual normalized difference vegetation index, as covariates. ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ Fine particulate matter ≤2.5 μm and nitrogen dioxide data sets were aggregated to the community area level by calendar year and merged into the yearly, monthly, and daily analytic panels by year. Normalized difference vegetation index values were derived from satellite‐based greenness estimates and averaged annually for each Chicago community area. Before inclusion, all 3 variables were standardized (Z‐scored) within each panel to facilitate model convergence and comparability of effect sizes. Normalized difference vegetation index was also joined to daily panels by community and year to ensure each observation was matched to the appropriate annual greenness level.

Death certificates for Chicago decedents are maintained by the Illinois Department of Public Health and provided by the Chicago Department of Public Health. In addition to decedent demographic characteristics like sex, race, date of death, and date of birth, we retrieved the geospatial markers of decedent residential address and community area. Each decedent had 1 primary and up to 3 contributing causes of death with their respective International Classification of Diseases, Tenth Revision (ICD‐10) codes listed on their death certificate. A CVD cause of death was classified as such if any primary or contributory causes of death contained an ICD‐10 code between I00 and I99. The following CVD subtypes were classified via ICD‐10 codes as coronary heart disease (CHD) I20‐I25, stroke I60‐I69, and myocardial infarction (MI) I21‐I22, per American Heart Association guidelines. ¹⁹

CAPRICORN (Chicago Area Patient‐Centered Outcomes Research Network) is a collaborative network of health care institutions and community organizations that leverages electronic health record data from participating sites as part of the larger Patient‐Centered Outcomes Research Network framework. CAPRICORN has a patient population of nearly 13 million and has been extensively used in prior health outcomes research. ²⁰ , ²¹ We queried all ED encounters across all participating CAPRICORN institutions from 2011 to 2022, which included encounter‐specific ICD‐10 diagnosis codes, patient demographic information, and patient residential address census tract. We joined each ED encounter to its respective community area using a spatial join of the patient’s residential address census tract. CVD and CVD subtype‐related ED encounters were identified using the same ICD‐10 codes previously described for CVD mortality.

Statistical Analysis

We selected 4 different temporal scales for this study: warm months, monthly, daily, and daily lag (3‐day rolling average). We aggregated CVD mortalities and ED encounters by community area and calculated the respective rate per 100 000 population by temporal scale (warm months, monthly, daily, and daily lag). Mean Tmax and relative humidity were also aggregated by community area and temporal scale. Warm‐months ACS estimates were joined by year and community area with the assumption that these estimates remained the same for each respective year.

Generalized Additive Model Heat‐Related Excess Morbidity and Mortality Estimates

The generalized additive model (GAM) is a robust framework for regressing nonlinear predictors against health outcomes and numerous studies have used GAMs for environmental health impact research. ²² , ²³ , ²⁴ It is well established that the relationship between temperature and CVD outcomes is nonlinear with a “U” shaped curve, which can be modeled in a variety of ways. ²⁵ , ²⁶ , ²⁷ GAMs can accommodate different temporal scales and lag structures, handle nonlinear multidimensional interactions with tensor smooths, and have improved interpretability over methods like distributed lag nonlinear models. ²⁸ , ²⁹

We selected final GAM covariates based on Akaike information criterion and Bayesian information criterion values. Final covariates included relative humidity, median age, median income, education level, racial composition, employment status, sex distribution, and total population (offset). The primary exposure was Tmax. Overdispersion was measured for each model using the deviance divided by the residual degrees of freedom with a value ≥1 indicating overdispersion. We employed 10‐fold cross‐validation for 4 GAMs addressing different cumulative heat exposure temporal scales: warm months, monthly, daily, and daily lag. All GAMs had a negative binomial link function to address overdispersion in the CVD outcome rates.

To account for seasonality and long‐term temporal trends, cubic regression spline functions for the year (warm‐months models), month (monthly models), or date (daily and daily lag models) were included in each model. Tmax was included as a penalized spline term function, with a prespecified basis dimension (k between 6 and 10 depending on the time scale: warm months, monthly, daily, and 3‐day moving average). Knot locations were not chosen manually; instead, knots were placed automatically based on the empirical distribution of the covariate, and the amount of smoothing was selected by restricted maximum likelihood with an additional shrinkage penalty. Continuous covariates with nonlinear effects (median age, relative humidity, and calendar time) were also modeled using penalized spline terms. Median age and relative humidity were modeled with penalized thin plate regression splines with small basis dimensions (typically k≈5), and calendar time (year in yearly models, numeric date in daily and monthly models) was modeled using cubic regression splines with shrinkage or to capture long‐term and seasonal trends. All other covariates (sex, race, ethnicity, education, income, employment, and air‐pollution and greenness indicators) were entered as linear terms because these variables represent long‐term background exposure rather than acute daily fluctuations.

The GAMs provided estimated CVD mortality or ED visit (morbidity) rates per 100 000 population for warm months, monthly, daily, and daily lag temporal scales by community area. Using the entire range of temperatures within each respective temporal scale, we identified the “peak threshold” (the temperature at which heat‐related excess CVD mortality or ED visit rates are maximized across all 77 community areas) and “highest positive threshold” (the highest temperature threshold that still results in positive heat‐related excess CVD mortality or ED visit rates across all 77 community areas). To calculate heat‐related excess CVD mortality or ED visit rates by community area, the cumulative below‐threshold mean estimated CVD mortality or ED visit rate was subtracted from the cumulative above‐threshold mean estimated CVD mortality or ED visit rate. To quantify uncertainty, we used bootstrapping (1000 samples) to estimate 95% CIs for heat‐related CVD mortality and ED visit rates at each temperature threshold. These statistical methods also help to prevent overfitting of the temperature–CVD relationship especially with regard to nonlinear terms.

Principal Component Analysis Clustering of Chicago Community Areas

To characterize multidimensional community vulnerability related to heat exposure, sociodemographic composition, green space, and cardiovascular health, we performed a principal component analysis (PCA) on standardized, community‐level features. Input variables included mean warm‐season maximum temperature, mean relative humidity, warm months normalized difference vegetation index, nitrogen dioxide, and fine particulate matter ≤2.5 μm, and demographic indicators from the ACS (median age, median income, percentage college educated, percent agewith high‐school education only, percentage employed, percentage unemployed, and race and ethnicity composition [% White, Black, Hispanic, Asian]).

Before PCA, all variables were centered and scaled. PCA was performed using singular value decomposition. Component loadings were examined to identify dominant gradients represented by each principal component. Clusters of communities were then identified by applying k‐means (k=3, chosen by the elbow method and inspection of PC1–PC2 separation) to the first 2 principal components. The derived clusters were then compared based on their respective mean CVD mortality and ED visit rates during warm months.

Sensitivity Analysis

Because mortality patterns changed during the COVID‐19 pandemic, we conducted a sensitivity analysis restricting the data to the prepandemic period (2011–2019) and we refit our warm‐months GAMs using the same model specification. This analysis evaluated whether the temperature–CVD relationship and estimated heat‐related excess morbidity and mortality were robust to removal of the COVID‐19 period.

We performed all geospatial and statistical analyses using R (R Core Development Team 2024, v4.3.2). Geoanalytic and mapping packages included “daymetr,” “exactextractr,” “sf,” “terra,” and “tmap.” The package “tidycensus” queried ACS tables and variables, “mgcv” provided the GAM functions, while “dplyr,” “ggplot2,” “lubridate,” “zoo,” “boot,” “cluster,” “factoextra,” and “caret” were used for data processing, analysis, and figure creation.

RESULTS

Study Cohort

From May to September 2011 to 2022, there were an initial 96 791 adult decedents and 782 416 ED visits within 325 530 unique adult patients with a Chicago address at the time of event. This yielded a total of 46 843 (48.4%) decedents with a CVD cause of death (50.9% male, 49.1% female; mean±SD age: 73.7 years [15.6 years]) and 219 044 (28.0%) total ED visits with a CVD diagnosis (44.5% male, 55.5% female; mean±SD age: 59.4 years [16.9 years]) during the study period, all of which were included in the rest of the analysis. The inclusion flow chart can be found in Figure S1.

Among CVD deaths, 33.3% were CHD deaths, followed by stroke (15.9%), and MI (10.7%). Among CVD ED visits, 17.1% were CHD ED visits, followed by stroke (8.9%), and MI (2.1%). The mean±SD May to September temperature was 26.07 °C (0.73 °C) and the mean±SD relative humidity was 67.7% (1.60%) on the day of death or ED visit. Table 1 provides the demographic characteristics of the CVD cohorts and Figure 1 displays the spatial distributions of temperature, relative humidity, CVD deaths, and CVD ED visits. Figures S2 through S4 display the spatial distribution of the environmental covariates.

Table 1.

Total Number of Adult CVD Deaths and ED Visits Across All Chicago Community Areas During Warm Months (May–September) From 2011 to 2022

Category	CVD deaths (n=46 843)	CVD ED visits (n=219 044)	Median (IQR) CVD deaths across all community areas	Median (IQR) CVD ED visits across all community areas
Race or ethnicity
Asian	1376 (2.9%)	3370 (1.5%)	7 (2–25)	19 (5–58)
Black Hispanic	318 (0.7%)	792 (0.4%)	4 (2–7.25)	8.5 (3–16)
Black non‐Hispanic	22 985 (49.1%)	105 073 (48%)	126.5 (26.25–429)	566 (213–1317)
Latino/a	879 (1.9%)	19 833 (9.1%)	7 (2–20)	124 (52–301)
Other	328 (0.7%)	59 341 (27.1%)	3 (1–7)	515 (243–806)
White Hispanic	4890 (10.4%)	13 066 (6%)	34.5 (5–92.25)	72 (24–202)
White non‐Hispanic	16 067 (34.3%)	17 569 (8%)	124.5 (18.75–301)	143 (62–316)
Age
65 y and above	34 148 (72.9%)	83 582 (38.2%)	365 (250–573)	675 (367–1177)
Below 65 y	12 682 (27.1%)	135 462 (61.8%)	124 (72–220)	1185 (556–1974)
Sex
Female	22 988 (49.1%)	121 639 (55.5%)	249 (158–400)	967 (514–1816)
Male	23 855 (50.9%)	97 375 (44.5%)	282 (161–432)	852 (448–1520)

CVD indicates cardiovascular disease; ED, emergency department; and IQR, interquartile range.

Figure 1. Mean maximum temperature (A), relative humidity (B), and warm‐months CVD morbidity (C) and mortality (D) rates per 100 000 population by Chicago community area from May to September (2011–2022).

CVD indicates cardiovascular disease; and ED, emergency department.

Generalized Additive Model Results

We estimated the association between temperature and cardiovascular outcomes using GAMs with cubic regression splines for Tmax, humidity, median age, and the temporal trend, while modeling all other community‐level sociodemographic and environmental characteristics as linear terms. Covariates included in the models and their distributions across warm‐season Tmax tertiles are shown in Table 2.

Table 2.

American Community Survey Estimates and Environmental Variables Across Chicago Community Areas Stratified by Maximum Temperature Tertiles Over the Study Period; Low (25.66 °C–25.99 °C), Medium (26.01 °C–26.16 °C), and High (26.17 °C–26.34 °C)

Variable	Category	Low Tmax tertile	Medium Tmax tertile	High Tmax tertile
	Total population	53 206.1	21 032.1	28 255.1
		(32 122.7–66 681.4)	(14 210.2–32 778.4)	(18 896.9–38 017)
	Median age, y	34.9 (32.5–38.1)	34.3 (31.6–37.3)	36.3 (32.3–39.7)
	Median income, $	64 687.4	33 661.5	41 928.6
		(54 541.3–92 740.4)	(28 336.2–50 022.8)	(33 791.5–53 292.8)
Education (%)	College	19.2 (14.9–27.5)	9.4 (6.1–13.3)	5.9 (4.3–9.1)
	High school	12.2 (8.1–15.8)	15.1 (11.9–18.3)	18.1 (15.7–20)
Race or ethnicity (%)	White	73.1 (62.4–80.2)	11.1 (4.2–44.5)	30.2 (2.8–57.2)
	Black	3.7 (2.2–10.2)	76 (23.2–92.2)	50.6 (5.5–94.6)
	Latino	19 (13.3–42.2)	4.9 (3–21.8)	9.3 (3.3–53.8)
	Asian	6.6 (4.5–11.2)	0.6 (0.3–10.5)	0.5 (0.3–1)
Employment (%)	Employed	53.7 (49.8–61.8)	38.9 (35–46.4)	39.7 (35.7–43.8)
	Unemployed	4 (3.4–4.5)	7.2 (5.9–8.9)	8.2 (5.6–9.6)
Sex (%)	Male	49.2 (48.3–50.4)	45.6 (43.5–47.8)	48 (45.4–49.5)
	Female	50.1 (49.2–50.9)	53.6 (51.8–55.4)	52 (50.1–54.3)
Environment	Nitrogen dioxide	16.9 (16.1–17.8)	16.6 (15.7–17.8)	15.6 (15.2–16.0)
	Fine particulate matter	9.0 (9.0–9.1)	9.1 (9.0–9.2)	9.2 (9.0–9.4)
	Normalized difference vegetation index	0.2 (0.2–0.2)	0.1 (0.1–0.2)	0.2 (0.2–0.3)

Tmax indicates maximum temperature.

Temperature Associations

Tmax demonstrated a significant, nonlinear association with both cardiovascular mortality and ED visits, with stronger and more consistent effects observed for mortality end points. In daily models, Tmax significantly predicted CVD ED visits (P=0.035) and was a strong predictor of CVD mortality (P=2.5×10⁻⁷) and CHD mortality (P=0.017). Temperature effects were weaker and nonsignificant for daily MI and stroke mortality. Humidity contributed independently to CVD mortality (P=0.046), suggesting that combined heat and humidity load may amplify cardiovascular risk.

GAM‐derived spline curvature demonstrated pronounced nonlinear temperature–mortality relationships (Figure S5). CVD and CHD mortality all showed substantial curvature at monthly and daily timescales (edf~2–4), indicating steep increases in mortality risk at higher temperatures. In contrast, temperature–ED visit relationships showed little to no curvature across subtypes or timescales (edf~0–1), except for warm months, which had a high degree of curvature.

Sociodemographic and Structural Predictors

Community‐level sociodemographic composition was strongly associated with cardiovascular ED visits and mortality. In ED models, higher proportions of Asian, Black, and Hispanic residents; lower median income; higher high‐school–only education; and higher unemployment were consistently associated with elevated ED visit rates across all cardiovascular subtypes (many P‐values <0.001). Sex composition was also influential, with higher proportions of female residents associated with lower ED use and higher proportions of male residents associated with higher rates. Mortality models reflected similar sociodemographic patterns, though with smaller effect sizes relative to ED encounters. Lower median income, higher unemployment, and higher prevalence of high‐school–only education were associated with increased CHD, CVD, and MI mortality. Racial and ethnic patterns differed somewhat between mortality and ED end points: Hispanic and Asian populations exhibited lower CHD and MI mortality risk, whereas Black population proportion showed mixed associations across mortality models. All significant covariates for the daily models can be found in Table S1.

Temperature Thresholds and Exposure–Response Curves

Across cardiovascular outcomes, GAM‐derived exposure–response curves revealed substantial heterogeneity in heat sensitivity by subtype and temporal scale (Table 3). For ED visits, peak excess rates were generally small and often nonsignificant outside of the warm‐season models. CHD ED visits peaked at 25.6 °C during warm months (−3.40 per 100 000; P<0.001) but showed no significant associations at daily (0.027 per 100 000 at 19 °C; P=0.19), lagged daily (0.300 per 100 000 at 10.5 °C; P=0.171), or monthly (2.56 per 100 000; P=0.243) timescales. CVD ED visits demonstrated the clearest heat sensitivity, with significant peaks in the daily (0.185 per 100 000 at 19 °C; P=0.035) and lagged daily models (1.85 per 100 000 at 10.5 °C; P=0.011), whereas MI and stroke ED visit peaks were not significant at any short‐term temporal scale. Collectively, ED outcomes exhibited weak and inconsistent acute heat effects, suggesting that short‐term temperature increases play a limited role in driving ED use.

Table 3.

Average Heat‐Related Excess Cardiovascular ED and Mortality Rates Across All 77 Chicago Community Areas at Their Respective Calculated Peak Excess Thresholds From May to September (2011–2022)

Outcome type	CVD subtype	Temporal scale	Peak Tmax threshold (°C)	Peak excess rate per 100 000 (SD)	P value
Emergency department visits	Coronary heart disease	Warm months	25.6	−3.398 (25.473)	<0.001
		Daily	19	0.027 (0.021)	0.19
		Daily (lag 0–3)	10.5	0.300 (0.158)	0.171
		Monthly	20.25	2.557 (2.427)	0.243
	All cardiovascular disease	Warm months	25.6	−10.027 (140.758)	<0.001
		Daily	19	0.185 (0.138)	0.035
		Daily (lag 0–3)	10.5	1.848 (0.933)	0.011
		Monthly	20.25	11.225 (12.646)	0.221
	Myocardial infarction	Warm months	25.6	3.674 (3.375)	<0.001
		Daily	10	0.005 (0.006)	0.723
		Daily (lag 0–3)	10.5	0.055 (0.026)	0.579
		Monthly	19.25	0.262 (0.387)	0.599
	Stroke	Warm months	25.75	9.051 (9.478)	<0.001
		Daily	22.5	0.014 (0.008)	0.3
		Daily (lag 0–3)	10.5	0.248 (0.105)	0.112
		Monthly	20.25	1.823 (1.371)	0.088
Mortality	Coronary heart disease	Warm months	27.75	9.193 (1.537)	0.004
		Daily	38.5	0.096 (0.057)	0.017
		Daily (lag 0–3)	38	0.097 (0.057)	0.039
		Monthly	30.75	2.837 (1.746)	0.002
	All cardiovascular disease	Warm months	25.65	20.705 (20.349)	<0.001
		Daily	39.5	0.048 (0.068)	<0.001
		Daily (lag 0–3)	38	0.049 (0.133)	<0.001
		Monthly	30.25	1.991 (2.146)	<0.001
	Myocardial infarction	Warm months	27.75	0.986 (0.293)	0.626
		Daily	39.5	0.004 (0.009)	0.341
		Daily (lag 0–3)	30.5	−0.005 (0.002)	0.057
		Monthly	30	0.234 (0.160)	0.403
	Stroke	Warm months	25.65	4.244 (3.193)	0.095
		Daily	24	0.001 (0.002)	0.225
		Daily (lag 0–3)	10.5	0.021 (0.024)	0.194
		Monthly	28	0.092 (0.161)	0.08

CVD indicates cardiovascular disease; ED, emergency department; and Tmax, maximum temperature.

By contrast, cardiovascular mortality displayed strong and consistent heat sensitivity across multiple exposure windows. CVD mortality showed significant peaks during warm months (20.71 per 100 000 at 25.65 °C; P<0.001), daily (0.048 per 100 000 at 39.5 °C; P<0.001), lagged daily (0.049 per 100 000 at 38 °C; P<0.001), and monthly periods (1.99 per 100 000 at 30.25 °C; P<0.001). CHD mortality similarly peaked at 27.75 °C during warm months (9.19 per 100 000; P=0.004), at 38.5 °C for daily exposures (0.096 per 100 000; P=0.017), and at 30.75 °C for monthly exposures (2.84 per 100 000; P=0.002), with lagged daily exposures yielding a significant peak (0.097 per 100 000 at 38 °C; P=0.039). Heat effects for MI and stroke mortality were weaker and largely nonsignificant; for example, MI mortality daily peaks were minimal (0.004 per 100 000 at 39.5 °C; P=0.341), and stroke mortality daily peaks were small (0.001 per 100 000 at 24 °C; P=0.225). Temperature response curves for ED visits and mortality can be found in Figures 2 and 3.

Figure 2. Estimated heat‐associated excess cardiovascular mortality across Chicago community areas, 2010 to 2022.

Modeled relationships between daily maximum temperature and average heat‐associated excess deaths per 100 000 population are shown for CHD (A), all CVD (B), MI (C), and stroke (D). Curves represent predictions from the warm‐months quasi‐Poisson GAM framework, with shaded areas indicating 95% CIs. Vertical dotted lines denote the temperature at which maximum excess mortality occurs (“peak threshold”), and vertical dashed lines indicate the highest temperature yielding >0 excess mortality. CHD indicates coronary heart disease; CVD, cardiovascular disease; GAM, generalized additive model; MI, myocardial infarction; and Tmax, maximum temperature.

Figure 3. Estimated heat‐associated excess cardiovascular ED visits across Chicago community areas, 2011 to 2022.

Modeled relationships between daily maximum temperature and average heat‐associated excess ED visits per 100 000 population are shown for CHD (A), all CVD (B), MI (C), and stroke (D). Curves represent predictions from the warm months quasi‐Poisson GAM framework, with shaded areas indicating 95% CIs. Vertical dotted lines denote the temperature at which maximum excess morbidity occurs (“peak threshold”), and vertical dashed lines indicate the highest temperature yielding >0 excess morbidity. CHD indicates coronary heart disease; CVD, cardiovascular disease; ED, emergency department; GAM, generalized additive model; MI, myocardial infarction; and Tmax, maximum temperature.

Principal Component Analysis Clustering for Vulnerability

PCA revealed 2 dominant gradients that together explained much of the variance in community‐level vulnerability features (Table S2; Figure S6). PC1 (socioeconomic disadvantage and heat exposure; 41.2% of variance) loaded positively on unemployment, percentage of Black residents, and higher warm‐season temperatures, and negatively on median income, college education, employment, and percentage of White residents. Higher PC1 scores therefore correspond to structurally disadvantaged, lower‐income, predominantly Black communities that also experience greater thermal burden. The geographic distribution is found in Figure 4.

Figure 4. Spatial distribution of community vulnerability clusters across Chicago.

Map of Chicago’s 77 community areas grouped into 3 vulnerability clusters derived from a principal component analysis of multidomain features, including heat exposure (daily maximum temperature and humidity), sociodemographic composition (income, education, employment, race, ethnicity), green space, and air pollution. Each color represents a distinct vulnerability profile, reflecting shared structural, environmental, and demographic characteristics. NDVI indicates normalized difference vegetation index; and PCA, principal component analysis.

PC2 (racial and ethnic composition and humidity contrast; 19.7% of variance) captured an orthogonal gradient separating predominantly Hispanic communities (large negative loadings) from communities with higher humidity and larger proportions of Black and Asian residents. PC2 therefore reflects demographic and microclimate contrasts independent of socioeconomic status. Together, these 2 components were used to cluster community areas into 3 distinct vulnerability profiles. These clusters had distinct CVD risk profiles for both morbidity and mortality (Figure 5; Figure S7).

Figure 5. Community vulnerability structure and its association with cardiovascular outcomes in Chicago.

A, Principal component analysis of community‐level vulnerability features, showing PC1 vs PC2. Points represent Chicago community areas and are colored by k‐means–derived vulnerability clusters. B, Distribution of average warm‐season cardiovascular disease emergency department visit rates and mortality rates (per 100 000 population) from 2011 to 2022 across the vulnerability clusters. CVD indicates cardiovascular disease; ED, emergency department; and PCA, principal component analysis.

Restricting the analysis to the pre‐COVID period produced results that were highly consistent with the primary models. Estimated temperature–mortality curves, peak heat thresholds, and predicted excess mortality changed minimally when excluding 2020 to 2022, indicating that the observed associations were not driven by pandemic‐related mortality anomalies (Figure S8).

DISCUSSION

In this large, community‐level analysis encompassing all 77 Chicago community areas from 2011 to 2022, higher temperatures were consistently and significantly associated with increased heat‐related excess CVD and CHD mortality across warm‐season, monthly, and daily timescales. Peak daily risk occurred at temperatures well above seasonal norms (eg, ~38 °C for CVD and CHD), with corresponding excess mortality rates of 0.026 and 0.098 deaths per 100 000 people per day, respectively. The temperature–mortality relationship was strongly nonlinear, with pronounced curvature for CVD and CHD across multiple timescales, indicating sharply increasing mortality risk at higher temperatures. In contrast, no significant temperature associations were observed for MI or stroke mortality, and ED visit rates showed little to no temperature responsiveness, suggesting that heat may disproportionately affect individuals at highest risk of fatal, rather than nonfatal, cardiovascular events.

Several community sociodemographic characteristics emerged as the most influential predictors of heat‐related cardiovascular mortality, and these patterns were strongly aligned with our PCA‐derived vulnerability structure. Communities with lower median income, higher unemployment, greater proportions of Black residents, and lower educational attainment consistently clustered at the high‐vulnerability end of the PCA spectrum (PC1), with loadings indicating that unemployment (loading=+0.37), lower income (−0.35), lower college attainment (−0.36), and higher Black population share (+0.33) were among the dominant contributors to structural disadvantage. These same characteristics were also the strongest linear predictors of increased CVD and CHD mortality in the GAM models, with significant effect estimates across daily, monthly, and warm‐season scales. Communities with higher proportions of Hispanic or Asian residents tended to fall into lower‐vulnerability PCA clusters and exhibited significantly lower heat‐related excess mortality. Median age also contributed meaningfully to vulnerability (PC2 loading=+0.09), and older communities showed higher excess mortality risk.

Although it is generally known that increased temperatures are associated with adverse cardiovascular outcomes, this area of research is relatively nascent, and the relationship is not yet well understood. Most exposure–response studies of this type have been performed in the past 5 years, with research focusing on CVD comprising <25% of the literature despite being the leading cause of death in many countries. ²⁶ Most studies have assessed the impact of extreme temperatures on CVD on a daily or monthly temporal scale over an extended longitudinal period and found that heat exposure increases the burden of heat‐related excess mortality and, in some cases, morbidity. ⁶ , ¹⁰ , ¹¹ , ¹² , ³⁰ , ³¹ There is rationale for studying longitudinal trends in CVD as long‐term exposures and behaviors influence acute event onset, however it is unclear how the estimated disease burden shifts by changing the window of exposure. ³² , ³³ A study from 2020 estimated that an average of 5608 annual deaths were due to heat (~2500 of which can be attributed to CVD) in a subset of US counties. ³⁴ One study estimated that there were between 5958 and 7144 excess CVD deaths associated with extreme heat in the contiguous United States during their 10‐year study period using monthly exposure‐response windows. ⁵ , ¹² Another study used a daily exposure‐response framework and found a 1.8% increase in daily CVD mortality for every 1 °C increase in Tmax above a predefined regional threshold. ³⁵ , ³⁶ Although our study is not directly comparable because we estimated heat‐related excess cardiovascular mortality using peak temperature thresholds, we did find that the heat‐related excess CVD mortality rate within Chicago was significantly related to warm‐month, monthly, and daily temperatures.

The primary focus of this study was to take an inductive approach to develop CVD‐specific thresholds for Chicago for various temporal scales. With these absolute thresholds we were able to estimate absolute values for excess cardiovascular morbidity and mortality attributed to heat exposure. This approach is different than much of the prior literature that applies a deductive approach using preset temperature thresholds to provide a risk value relative to another preset threshold. ³⁷ , ³⁸ A deductive approach can be useful, but it also can miss critical temperature thresholds that are disease specific while also providing relative values that may not be applicable to stakeholders. Our study identified specific points along the temperature spectrum most predictive of increased rates of adverse cardiovascular outcomes and applied those thresholds to estimate the corresponding heat‐related excess cardiovascular morbidity and mortality at the community area level. Assuming a population of 3 million residents, we found that on days >38 °C, ~2.6 CHD deaths in Chicago are heat related, and for months with an average Tmax of 30 °C, ~28 CVD deaths in Chicago are heat related. These absolute values provide utility to population health stakeholders for vulnerability indexing and resource allocation.

The association between heat and specific subtypes of CVD, namely stroke and MI, remains unclear. In studies evaluating the effect of extreme heat on MI incidence, approximately half found a significant association and half did not. ³² , ³⁴ Research on stroke outcomes is similarly complicated. ³⁹ Our study found that there were no significant associations between either MI or stroke excess mortality or ED visit rates and increased temperatures. Furthermore, our study found no evidence of increases of heat‐related excess cardiovascular ED visits due to temperature, which is in line with the existing literature. ³⁵ , ³⁷ , ⁴⁰ Despite plausible biological mechanisms, there is inconsistent evidence that heat and cardiovascular morbidity are related. However, we did find spatial patterns that seemed to suggest that the community areas with the highest heat‐related excess CVD mortality rate usually did not overlap with community areas with the highest heat‐related excess CVD ED visit rate. This indicates that there may be underlying features not captured by our models that differentiate these 2 cardiovascular outcomes.

A limitation of this study is the assumption that patient or decedent residential address accurately depicts the heat exposure for a given health event. Although it is possible that the home address at the time of event is not representative of the burden of heat exposure, we believe that it serves as the best method for geocoding despite its potential shortfalls. Likewise, if a patient or decedent did not have an address, they could not be geocoded, which potentially excluded some of the populations most vulnerable to heat. Another limitation is the representativeness of the ED data. CAPRICORN is the most comprehensive data‐sharing network of medical institutions in Chicago; however, it does not encompass all health care sites in the city. This underrepresentation likely contributes to a bias toward the null for ED visits, particularly on the South and far North sides of Chicago. Though statistical methods to improve representation were considered, we determined that imputation methods involved assumptions that were too strong. This suggests that the true impact of temperature on ED visits may be stronger than what was observed in our results. Another limitation is the potential for overestimation of events due to CVD by not exclusively using the primary diagnosis code; however, we believe that such a restriction would remove the multifaceted nature of heat‐related outcomes. There is also the potential for statistical model overfitting and Type 1 error rate, though steps were taken to mitigate these issues. Furthermore, we recognize that the clusters may be specific to this data set and time period and that external validation would be an important next step. Future work should seek to understand some of the underlying clinical drivers of cardiovascular morbidity by developing robust computational phenotypes of CVD using longitudinal electronic health records.

Conclusions

In conclusion, this study provides a novel approach to quantifying heat‐related cardiovascular risk by integrating heat‐related excess morbidity and mortality estimates across daily, monthly, and warm‐month temporal scales while accounting for both short‐term and cumulative heat exposure. By clustering community areas based on shared vulnerability profiles, our method offers a more granular understanding of how extreme heat affects cardiovascular health during short‐ and long‐ term heat exposure. This framework can be replicated in other urban settings to identify high‐risk communities and inform targeted interventions. The identification of sociodemographic risk factors, such as education level, income, and racial and ethnic composition, underscores the importance of addressing social determinants of health in climate adaptation efforts. These results can help guide heat early warning systems, urban planning decisions, and resource allocation to reduce the burden of heat‐related cardiovascular mortality. As extreme heat events become more frequent and severe due to climate change, refining risk assessment methods and improving public health interventions will be essential in protecting vulnerable populations.

Sources of Funding

This work was made possible with funding from the American Heart Association Predoctoral Fellowship 24PRE1193628 awarded to Peter M. Graffy as well as support from the Health Sciences Integrated Program (HSIP) as part of the Institute for Public Health and Medicine (at Northwestern University.

Disclosures

No authors have any conflicts of interest to disclose.

Supporting information

Tables S1–S2

Figures S1–S8