Discrepancies Between Personal and Ambient Temperatures at Hourly Scale: Insights from Wearable Sensors and Environmental Modeling

Abstract

Accurate assessment of personal exposure to ambient temperature is essential for understanding temperature-related health impacts. However, the bias from using ambient temperature estimates as proxies for personal exposure remains underexplored, particularly at the hourly scale. This study integrated wearable temperature sensors with a high-resolution (1 km, hourly) ambient temperature model to evaluate individual-level ambient exposure among 94 individuals in Connecticut, USA, monitored across seasons between October 2023 and January 2025. Personal temperature was consistently higher than ambient temperature, with a larger difference during cooler months. The hourly difference exhibited a distinct unimodal diurnal pattern, smallest in the early afternoon. Linear mixed-effects models identified ambient temperature, hour of day, month, solar radiation, and nightlight index as the key predictors of personal temperature and its deviation from ambient conditions. The temperature difference was well characterized by the model (marginal R² of 0.854). Extreme gradient boosting with SHapley Additive Explanations confirmed ambient temperature and hour of day as the most influential features, with albedo and other environmental factors showing smaller effects. Findings highlight systematic seasonal and diurnal biases in ambient-based metrics, underscoring the need to account for these patterns when assessing thermal exposures.

Graphical Abstract

Introduction

The thermal threats posed by global climate change are increasingly intruding into daily life.^{1, 2} In 2019, 1.69 million deaths globally were attributable to daily non-optimal (heat and cold) temperatures.³ Even acute exposure may induce elevated blood pressure⁴ and an increased risk of myocardial infarction⁵. Despite increasing recognition of the health risks, most studies use ambient temperature data from weather stations or spatiotemporal models as proxies for personal exposure.^{6, 7} This raises concerns, as individuals today spend the majority of their time indoors, where temperatures can differ substantially from outdoor conditions,^8–11 potentially introducing substantial exposure misclassification. It is commonly assumed to be random, biasing estimates toward the null.¹⁰ However, this remains debatable, as the ability to modify personal exposure varies across individuals and sociodemographic groups, affecting their exposure to indoor and outdoor temperatures.

Understanding the temporal patterns and determinants of the temperature discrepancy between personal exposure and ambient level is critical for evaluating the direction and magnitude of exposure bias, which may distort exposure–response relationships and misguide public health interventions. Yet, only a few studies have preliminarily investigated personal temperature exposure,^12–14 and even fewer assessed its hourly-level discrepancy from ambient conditions. Occupational studies show that outdoor workers often experience higher personal temperatures or heat index stress than ambient data suggest.^15–17 Other studies, mostly conducted during summer or extreme heat events, report personal temperatures exceeding ambient levels.^18–20 However, it remains unclear whether such discrepancies persist across seasons or vary diurnally, especially among vulnerable populations.

To address this gap, we conducted a prospective panel study among adults with type 2 diabetes in Connecticut, collecting personal temperature exposure data alongside ambient temperature estimates at an hourly resolution. This study aimed to characterize temporal patterns and determinants of discrepancies between personal and ambient temperature exposures. By systematically examining these differences, we provide evidence to inform accurate exposure assessment and support the development of context-specific strategies for temperature-related health protection.

Materials and Methods

Study population and data sources

This study was conducted as part of a prospective observational research which aims to investigate environmental and behavioral risk factors associated with glycemic variability among individuals with type 2 diabetes. As of January 2025, temperature measurements have been completed for 94 participants who lived across the state of Connecticut, USA (Figure S1). Details of the parent study design and participant recruitment process have been published.²¹

To systematically identify the key determinants of personal temperature exposure, we integrated variables encompassing sociodemographic characteristics, residential environments (Table S1), and census-level neighborhood conditions (Table S2). Demographic information were collected at baseline via a structured questionnaire. Participants’ residential addresses were geocoded and linked to high-resolution meteorological and environmental datasets, including hourly meteorological variables from the ERA5-Land reanalysis dataset,²² greenness represented by the monthly normalized difference vegetation index (NDVI),²³ annual land imperviousness,²⁴ and nighttime light intensity.²⁵ Socioeconomic conditions were characterized using 2018–2022 5-year estimates from the American Community Survey.²⁶ Census data were accessed via the tidycensus R package (version 1.7.1) at the census block group or tract level, depending on availability.²⁷

Personal exposure measurement

Ninety-four participants were equipped with a wearable temperature logger (DROP D2HS Heat Stress Monitor, Kestrel Instruments^®) to continuously record personal dry-bulb temperature for two weeks at a 10-minute logging interval from October 2023 to January 2025. Participants were scheduled for monitoring periods based on rolling recruitment and individual availability. While the number of participants per month varied, enrollment was distributed across multiple seasons to capture seasonal variability at the population level. We used a total of 48 loggers, which were rotated across participants based on the monitoring schedule. Each device was factory-calibrated with certificants, confirming that its performance met NIST-traceable testing standards. As the temperature sensor is not field-adjustable and rarely drifts out of specification, no additional calibration or recalibration was performed during the study. Before each deployment, loggers were checked for battery level and physical integrity, and logging parameters were configured using the Kestrel LiNK mobile application. Participants were instructed to hang the logger from a personal item they carried consistently, such as a keychain, pocket, or purse, to keep the sensor exposed to ambient air. As such, the loggers were designed to capture ambient temperature in the participant’s immediate environment, rather than body temperature or activity-specific exposures. Participants were asked to keep the loggers near them, ideally wearing them as much as possible, but placing them nearby if uncomfortable, during sleeping, or showering. Data quality control is provided in Section S1.

Ambient temperature estimation

We derived 2-m dry-bulb temperature by a high-resolution (1 km, hourly) spatiotemporal model specifically for Connecticut, following the approach described in our previous work.^{28, 29} The model was trained on hourly observations from 10 National Oceanic and Atmospheric Administration land weather stations³⁰ and incorporated predictors including Moderate Resolution Imaging Spectroradiometer land surface temperature,³¹ ERA5-Land reanalysis temperature,²² elevation,³² land use type,³³ vegetation cover (measured by NDVI), geographic coordinates, day of year, and year. Details of the modeling method and predictor variables are described in Section S2 and Table S3. To derive hourly estimates, daily maximum and minimum predictions were first generated using random forest models and then temporally disaggregated using hour-specific fixed-effects models aligned with ERA5-Land diurnal patterns. The final hourly estimates showed high agreement with ground measurements (overall R² of 0.964 and root-mean-square error of 1.88°C; Table S4).

Statistical analysis

We calculated statistics and Pearson correlations for temperature metrics and visualized their distributions. To identify the diurnal patterns, we applied dynamic time warping (DTW), a time-series alignment algorithm that allows sequences with similar shapes but phase-shifted features to be compared meaningfully.³⁴ In our case, DTW aligned each diurnal temperature curve to others by allowing up to ±1 hour of temporal shift. This alignment produced a pairwise distance matrix that quantitatively reflected the shape similarity between diurnal temperature curves. We then used this matrix as input for K-means clustering to group diurnal patterns with similar temporal profiles. After testing up to eight clusters, we identified two distinct intra-day patterns based on within-cluster variability and practical relevance.

To examine the determinants of personal temperature and temperature difference, we employed linear mixed-effects models given the repeated-measures structure of the data. We accounted for within-individual correlation by including a random intercept for each participant. Fixed effects were drawn from three dimensions of spatiotemporal variables regarding personal-level sociodemographic characteristics and residential environment, and census-level socioeconomic indicators. To capture temporal patterns, we additionally included several time-related variables: hour of day, weekday/weekend, and month of year. Hour of day and month of year were modeled using natural splines with 4 degrees of freedom, while the other variables were modeled as linear predictors. To construct the model with the optimal relationship with geographic variables, we selected the buffer size with the biggest absolute standardized coefficients of univariate regression (Table S5).

To further explore potential nonlinear and interaction effects between predictors and hourly temperature exposure, we applied the Extreme Gradient Boosting (XGBoost) algorithm.³⁵ We then used SHapley Additive exPlanations (SHAP) values to assess the feature importance of the XGBoost model.³⁶ SHAP values quantify the marginal contribution of each feature to individual predictions, providing interpretable explanations for complex machine-learning models.

All statistical analyses were conducted in R (Version 4.4.2), using the packages including dtwclust, lme4, splines, xgboost, SHAPforxgboost.

Results

Personal temperature exposure and its temporal patterns

A total of 94 participants were included, with a mean age of 58.8 years and a slight female predominance (57%). Most identified their race as white (64%), with approximately half reporting full-time employment and home ownership (Table S1). Participants resided in neighborhoods with relatively low population density (1.9 × 10³/km²) but high rates of housing occupancy and employment (Table S2). The surrounding built and natural environments were characterized by moderate vegetation cover (mean NDVI = 0.30), and substantial impervious surfaces (mean = 46.3%).

We observed substantial variability between personal, ambient, and difference (personal minus ambient) temperatures (Figure 1A–B; Table S6). Although the correlation between personal and ambient temperature was moderate (r = 0.46; Table S6), personal temperature was markedly higher than ambient temperature (averages of 22.1 °C vs. 12.4 °C), with a mean difference of 9.9 °C. Figure 1B shows the hourly time series of personal and ambient temperatures across all participants. The temperature difference was smaller during the warm season and more pronounced during the cold season. The observed breaks in the personal temperature series reflect calendar periods during which no participants were enrolled in the study. This substudy was conducted as part of a broader investigation involving continuous glucose monitoring (CGM). During the summer months, particularly August, some participants were unavailable due to vacation, and others were hesitant to wear the CGM device due to concerns about visibility and disclosing their diabetes status while wearing short sleeves. These gaps represent recruitment timing, not missing data from enrolled participants. Two distinct diurnal patterns of temperature difference were identified using DTW clustering (Figure 1C). Both followed a unimodal trough-like shape, reaching their minimum during early afternoon hours. As Figure 1D illustrates, cluster 1 (in red) primarily included measurements from the warm season and showed smaller temperature differences. In contrast, cluster 2 (in blue), mostly from the cold season, exhibited larger discrepancies.

Figure 1. Distributions and temporal patterns of personal, ambient, and difference temperatures.

(A) Boxplots of hourly personal, ambient, and difference (personal minus ambient) temperatures. (B) Time series of personal and ambient temperatures across participants for each hour from October 2023 to January 2025. (C) Diurnal patterns of difference temperature stratified by two clusters identified using dynamic time warping (DTW), shown as mean and 95% confidence intervals. (D) Scatterplot of ambient temperature and difference temperature, with color indicating DTW-derived clusters.

Determinants of personal temperature and its discrepancy with ambient temperature

We identified both temporal and spatial factors associated with hourly personal temperature and its deviation from ambient temperature (Table 1; Figure S2). Nonlinear effects of time-of-day and month were significant in both models. Personal temperature increased with ambient temperature (β = 0.14, p < 0.001), but the discrepancy between personal and ambient temperature narrowed significantly as ambient temperature rose (β = −0.85, p < 0.001). Participants living in rental apartments experienced significantly higher personal temperatures compared to homeowners. Several environmental predictors, including NDVI, nightlight intensity, albedo, and solar radiation, were negatively associated with personal temperature, suggesting cooling effects from vegetation, surface reflectivity, and behavioral adaptation to solar exposure. These variables remained statistically significant in the temperature difference model, although their effect sizes were small, suggesting limited but detectable roles in modulating personal-environment thermal discrepancies. The mixed-effects model explained 25.0% of the variance in personal temperature (marginal R²) and 85.4% of the variation in temperature difference (Table 1). We further stratified the linear mixed-effects model by season to explore whether the influence of individual-level predictors varied across warm and cold periods. As shown in Table S7, individual sociodemographic factors (e.g., housing type, marital status, employment, income) had stronger associations with personal temperature in the cold season than in the warm season.

Table 1.

Predictors of hourly personal temperature exposure and its difference with ambient temperature

	Personal temperature			Temperature difference
	Coefficient	95%CI	P value	Coefficient	95%CI	P value
Hour (Spline term 1 of 4)	0.68	(0.52, 0.84)	<0.001	0.58	(0.42, 0.74)	<0.001
Hour (Spline term 2 of 4)	2.14	(1.99, 2.30)	<0.001	1.99	(1.84, 2.14)	<0.001
Hour (Spline term 4 of 4)	0.81	(0.68, 0.94)	<0.001	0.72	(0.60, 0.85)	<0.001
Month (Spline term 1 of 4)	−1.38	(−2.38, −0.38)	0.007	−2.02	(−2.97, −1.07)	<0.001
Month (Spline term 2 of 4)	1.19	(0.20, 2.18)	0.019
Month (Spline term 3 of 4)	3.57	(1.55, 5.58)	<0.001	2.19	(0.28, 4.10)	0.026
Weekend (ref = weekday)	0.13	(0.07, 0.20)	<0.001	0.18	(0.12, 0.24)	<0.001
Ambient temperature, °C	0.14	(0.13, 0.15)	<0.001	−0.85	(−0.86, −0.84)	<0.001
Rental apartment (ref = own a house)				1.54	(0.09, 2.98)	0.041
NDVI, %	−0.01	(−0.01, −0.00)	0.025
Nightlight index, nW/cm3/sr	−0.01	(−0.02, −0.01)	<0.001	−0.01	(−0.02, −0.01)	<0.001
Albedo, %	−0.01	(−0.02, −0.00)	0.020
Solar radiation, *106, MJ/m2	−0.01	(−0.01, −0.00)	<0.001	−0.01	(−0.01, −0.00)	0.003
	Marginal R2 = 0.250; Conditional R2 = 0.507			Marginal R2 = 0.854; Conditional R2 = 0.900

Both models are based on data from 94 individuals and 27346 hourly observations. Spline terms for time variables (e.g., hour and month) were modeled using natural cubic splines with specified degrees of freedom.

We further applied SHAP analysis to quantify the relative importance of all predictors in the nonlinear XGBoost models for personal temperature and its difference with ambient temperature (Figure 2; Table S8). Ambient temperature was the strongest predictor in both models, followed by hour-of-day and albedo. Albedo consistently showed strong negative SHAP values, reinforcing its role in attenuating thermal exposure. For temperature difference, SHAP values further emphasized the dominant effect of ambient temperature and hour-of-day, consistent with the linear model. Other variables indicating vegetation, urbanization, and socioeconomic levels had smaller but consistent impacts, suggesting that both individual-level and environmental determinants shape thermal exposure disparities.

Figure 2. Variable importance for the personal temperature (left) and its deviation from ambient temperature (right).

Each dot represents an individual data point. The horizontal position indicates the SHAP value, which reflects the contribution of that feature to the model prediction (i.e., positive SHAP values increase the predicted outcome while negative values decrease it). Dot color represents the feature value, illustrating how different values of the variable influence the prediction. A wider horizontal spread indicates greater influence on the model output. Features are ranked by their mean absolute SHAP value (i.e., average importance across all observations). Factor names in green are related to environmental information, in orange related to census-level socioeconomic information, and in purple related to personal-level sociodemographic information.

Discussion

In this study, we found moderate correlations between personal temperature and ambient temperature with a consistent and systematic discrepancy between them (9.9°C). Several previous studies conducted during summer or heatwaves have reported smaller personal–ambient temperature differences, typically around 2–5°C.^{12, 19, 20} However, our study captured a substantially larger discrepancy, driven primarily by wintertime exposures in a well-heated U.S. population, underscoring the importance of characterizing thermal environments across seasons and heating contexts. A study conducted in China reported that the correlation between personal and ambient temperatures decreased from 0.56 in summer to 0.37 in winter.¹² Notably, our study provides an additional insight by characterizing personal–ambient temperature differences at an hourly scale. We observed a consistent diurnal pattern with smaller discrepancies in the afternoon and larger ones at night, regardless of season. In contrast, a study from India reported a reversed diurnal pattern in summer, where ambient temperatures (~35°C) exceeded personal exposure levels¹³. While this finding supports the existence of personal-ambient discrepancies, the direction and magnitude of the bias differed from our observations in Connecticut. This contrast highlights that such systematic discrepancies are dependent on regional climatic conditions and should be carefully evaluated in exposure assessment and health effect estimations.

While our model achieved a high R² of 0.833 in explaining the discrepancy between personal and ambient temperature at the hourly level, the R² for personal temperature was much lower. This reflects the relative stability of personal temperature and its limited responsiveness to ambient or contextual variables. In contrast, ambient temperature varied more substantially and largely drove the observed discrepancies. Previous studies also found it challenging to model personal temperature even including more variables concerning lifestyle, detailed time-activity, and building characteristics.^12,13 Additional studies have explored personal heat index exposure in specific subpopulations or intervention settings, highlighting the influence of microenvironments and behavioral patterns.^{37, 38} Although these studies did not directly compare personal and ambient temperatures, they further underscore the complexity of individual-level thermal exposure. Our findings demonstrate that the bias introduced by using ambient temperature as a proxy is systematic and can be modeled with high accuracy. This provides a practical framework for correcting exposure misclassification in health effect studies, particularly when high-resolution personal monitoring is not feasible.

This study identifies ambient temperature and temporal factors (e.g., hour of day, month of year) as primary determinants of personal temperature exposure. While discrepancy exists, ambient temperature largely defines the surrounding thermal condition.¹⁴ Temporal factors reflect complex interactions between solar cycles and human time-activity patterns, making them essential in modeling personal temperature. In line with the urban heat island mechanism, lower surface albedo, such as urban construction surfaces, was associated with elevated localized radiative heat load.³⁹ Other environmental factors, such as solar radiation, may influence personal temperature exposure differently than they affect ambient levels.

Individual behaviors may further modulate these effects based on perceived thermal conditions. Besides, previous studies conducted in the United States have reported strong associations between an individual’s disadvantaged socioeconomic status and increased personal heat exposure, particularly among low-income groups.^{40, 41} Our stratified results suggest that individual characteristics contributed more to variation in personal temperature during the cold season. These seasonal differences highlight the importance of considering context-specific vulnerability when interpreting personal exposure patterns, especially in colder months when indoor heating practices may vary substantially by socioeconomic status. In addition, the behavioral characteristics of individuals with type 2 diabetes may also influence the observed temperature patterns. Although we did not collect direct evidence of indoor or outdoor activity patterns in this study, prior research suggests that older adults and women, who comprised a large proportion of our sample, may be more likely to engage in a sedentary or indoor-based lifestyle.^42–44 These patterns may result from mobility limitations, comorbidities, or caregiving responsibilities. In addition, individuals with diabetes may intentionally avoid outdoor activity during periods of extreme weather to mitigate health risks.⁴⁵ These behavioral adaptations may contribute to lower variability in personal temperature and should be considered when interpreting the findings and assessing generalizability to other populations.

This study has several limitations. First, the sample’s geographic and clinical specificity may limit generalizability and some relevant covariates were unavailable. This study did not collect participants’ real-time location or detailed building characteristics, such as heating conditions. GPS tracking was considered but ultimately not implemented due to participant concerns regarding privacy and reluctance to share precise location information. While the sample included participants with varied sociodemographic and housing characteristics, including renters and individuals reporting financial difficulty, we did not directly assess energy insecurity or building-level insulation and heating conditions. Air conditioning was used by about half of participants, but adding this variable did not change model results, likely because the high prevalence of residential heating (>99%) and AC (>85%) in Connecticut reduces between-household variation in indoor climate;⁴⁶ nonetheless, such seasonal indoor climate control may still help interpret observed seasonal discrepancies. Second, this study monitored each participant for a 2-week period, with enrollment distributed across different seasons from October 2023 to January 2025. While this design enabled cross-sectional assessment of diurnal and seasonal variability in personal–ambient temperature discrepancies, it does not capture within-person seasonal adaptation or interannual variation. In addition, the short duration of monitoring per participant may not fully reflect habitual temperature exposures. In addition, the current study focused solely on temperature as the exposure metric. Future research could extend this work by integrating additional environmental dimensions, such as humidity or composite thermal indices, and combining them with physiological response data to better elucidate causal pathways linking thermal stress to health outcomes. Such efforts would provide a stronger scientific basis for developing personalized and targeted interventions against thermal-related health risks.

Although this study did not directly examine health outcomes, the observed discrepancy between personal and ambient temperature exposures has meaningful implications for chronic disease populations, particularly individuals with diabetes. Growing evidence links ambient temperature to both diabetes incidence and glycemic control.^{45, 47} Hyperglycemia can impair thermoregulation by reducing skin blood flow and sweating, and by promoting dehydration, which worsens insulin resistance and hepatic glucose production.^{48, 49} Diabetic individuals are particularly vulnerable to heat-related risks due to impaired vasodilation and coexisting cardiovascular or kidney disease.⁴⁵ Understanding what modulates personal temperature exposure could guide individualized or population-level strategies. For instance, reducing cold exposure during winter may be especially relevant for diabetic patients in resource-limited settings. Future studies could explore how thermal profiles interact with chronic disease status to shape health risks and adaptation needs.

Supplementary Material

Additional details on personal temperature data quality control and ambient temperature modeling methods; summary of demographic, socioeconomic, and environmental characteristics of the study population; list of predictor variables for high-resolution ambient temperature modeling; regression diagnostics and model performance evaluation; analysis of environmental variables across different spatial buffers; statistics and correlations for temperature metrics and covariates; season-stratified linear mixed-effects models; and spatial distribution of study participants and meteorological stations.

Synopsis.

This study identifies systematic seasonal and diurnal biases in personal versus ambient temperature, enhancing accuracy in climate–health exposure assessment.