Estimating facility-level greenhouse gas emissions in U.S. hospitals with machine learning imputation of incomplete self-reported data

Abstract

The U.S. healthcare industry has the largest per-capita carbon footprint compared with those of other countries. Despite growing commitments and efforts to decarbonization hospitals face challenges in reducing emissions in part because hospital-level estimates of greenhouse gas (GHG) emissions are limited or unavailable. This study estimates hospital-level GHG emissions from 38 sources across seven operational sectors for 283 hospitals, representing approximately 4% of U.S. hospitals in 2020, that voluntarily participated in the Practice Greenhealth Environmental Excellence Award program. Using our previously developed and validated machine learning–based imputation approach based on gradient boosting machines, this study addresses incomplete self-reported operational and activity data to enable large-scale facility-level emissions accounting. Total GHG emissions from these hospitals were estimated at 9.3 megatonnes (Mt) carbon dioxide (CO₂) equivalent, corresponding to nearly 2% of estimated U.S. healthcare emissions. Energy use, food services, and waste management were the largest contributors. While total emissions increased with hospital size, GHG emissions per staffed bed declined as hospital size increased, suggesting greater resource efficiency in larger facilities. Smaller hospitals exhibited both higher emissions intensity and a greater prevalence of missing activity data. These findings demonstrate that hospital-level GHG accounting can identify priority areas for targeted mitigation and reveal substantial variation in emissions across facilities, indicating opportunities for emissions reduction without compromising healthcare quality. The results underscore the need for standardized data collection to improve the accuracy of hospital-level GHG emissions estimates and to support scalable decarbonization strategies.

1. Introduction

Healthcare, which aims to improve public health by preventing disease, is a significant source of greenhouse gas (GHG) emissions. Healthcare is one of the largest service industries and is extremely carbon intensive (Eckelman et al., 2020; Hu et al., 2022; MacNeill et al., 2017; Pichler et al., 2019). According to multiple input–output–based estimates that account for both direct emissions from healthcare activities and indirect emissions along healthcare supply chains, healthcare-related consumption is associated with approximately 4.4%–4.6% of annual global net GHG emissions (HCWH, 2019; Lenzen et al., 2020; Watts et al., 2018). Compared with healthcare services in other countries, the carbon emission per capita from the US healthcare industry is the largest, but this has not translated into better quality of care (Eckelman et al., 2020; Pichler et al., 2019; Watts et al., 2019, 2021). Therefore, there are significant opportunities to reduce GHG emissions without compromising the quality of healthcare delivery, for example by reducing medical waste and avoiding overtreatment (Berwick and Hackbarth, 2012; Lenzen et al., 2020).

Globally, healthcare decarbonization has emerged as a critical component of climate mitigation and resilience planning (Leal Filho et al., 2024; Lenzen et al., 2020; WHO, 2025). Improved measurement of healthcare-related emissions is increasingly emphasized as an important foundation for implementing mitigation actions, including energy efficiency improvements, low-carbon procurement, sustainable food systems, transportation management, and waste reduction (Eckelman and Sherman, 2018; Tennison et al., 2021; Watts et al., 2021). In addition, emissions accounting can support adaptation and resilience planning by informing assessments of climate-sensitive resource dependencies, such as energy and water use, particularly in regions vulnerable to climate extremes (NAM, 2024).

In the U.S., leading organizations such as the US National Academy of Medicine, the Department of Health and Human Services (HHS), and the Joint Commission have developed initiatives to reduce greenhouse gas emissions in the healthcare industry (HHS, 2022; Joint Commission, 2022; NAM, 2022). In 2022, the HHS Health Sector Climate Pledge, a voluntary initiative that aims to reduce healthcare GHG emissions by 50% by 2030 and to achieve net-zero emissions by 2050, with approximately 15% of U.S. hospitals participating at the time (HHS, 2022; The White House, 2022). Although the pledge was later removed from the HHS website in 2025, this change does not reflect a withdrawal of hospital engagement in decarbonization efforts. Hospital participation has continued through non-governmental and sector-led initiatives with similar goals, including the Health Sector CARES Pledge led by Health Care Without Harm (“HCWH, 2022 Pledge,” 2025). This continued engagement underscores sustained efforts to reduce GHG emissions in the healthcare sector and reflects the urgency of addressing emissions to promote more sustainable and resilient health systems.

GHG emissions in healthcare are generally estimated using three main approaches. At the national or regional scale, environmentally extended input–output (EEIO) models are used to quantify emissions associated with healthcare consumption across supply chains (Eckelman et al., 2020; Eckelman and Sherman, 2018; Lenzen et al., 2020; Pichler et al., 2019; Wu, 2019). This top-down approach examines carbon emissions throughout the supply chain of products and services by applying environmental input-output datasets, using economic data, at a regional, or country level (Ingwersen et al., 2022). At the other end of the spectrum, life cycle assessment (LCA) studies estimate the climate impacts of specific medical procedures, technologies, or products. Examples include assessments of surgical procedures (Ferrero et al., 2022; Goel et al., 2021; MacNeill et al., 2017), magnetic resonance imaging (Leapman et al., 2023), hysterectomy (Thiel et al., 2015) and endoscopic procedures (López-Muñoz et al., 2023) and other medical activities (Chua et al., 2021; Gordon et al., 2021; Thiel et al., 2018). While LCA provides detailed, bottom-up insights into individual activities, extending this approach to an entire health system would require substantial data and resources due to the diversity and complexity of healthcare services. Hospital-level GHG accounting complements both EEIO benchmarks and procedure-level LCA evidence by integrating emissions across multiple sectors within individual facilities. Operating at the organizational scale, hospital-level estimates support accountability, operational decision-making, and translation of sector-wide targets into actionable mitigation strategies. Recent studies illustrate the value of this approach, including analyses of transport-related emissions at a hospital in Greece (Bozoudis and Sebos, 2021), financial-data–based emissions estimates from a German hospital (Keil, 2023), and survey-based emissions estimates from 33 hospitals in Switzerland (Keller et al., 2021). While these studies provide valuable insights for individual hospitals or specific clinical activities, large-scale analyses spanning multiple sectors and facilities remain limited. In particular, few studies systematically address the pervasive challenges of incomplete activity data in healthcare settings.

This paper estimates hospital-level GHG emissions across seven key operational sectors: energy, food, anesthetics, waste management, recycling, chemicals, transportation, and water consumption. These sectors capture major sources of hospital resource use, including direct on-site emissions (scope 1), emissions from purchased energy (scope 2), and selected operational scope 3 activities. However, they do not encompass the full healthcare carbon footprint, as several important scope 3 categories, such as pharmaceuticals, single-use medical devices, and other upstream supply-chain activities, are not included due to data limitations.

The analysis draws on a unique dataset collected by Practice Greenhealth, a consortium of hospitals and healthcare systems focused on reducing environmental impacts. The dataset documents operational activities for 283 hospitals in 2020, based on voluntary multiple-choice, open-ended, qualitative, and quantitative survey responses. A methodological framework previously developed by the authors, using gradient boosting machine–based imputation, was applied to address missing data in hospital-level consumption records (Yin et al., 2023). Building on this framework, this study provides hospital-level GHG emissions estimates from major sectoral sources, examines the influence of key predictors in imputing missing activity data, and evaluates how uncertainty in emissions estimates varies with data completeness. The analysis further explores how hospital carbon footprints vary by hospital size and between academic and non-academic facilities.

This work advances hospital-level GHG accounting in several important ways. First, this study integrates heterogeneous operational and activity data across seven major sectors, including energy, food, anesthetics, waste, chemicals, transportation, and water, into a unified bottom-up emissions estimation framework at the facility level. Second, the application demonstrates how machine learning–based imputation enables GHG accounting in real-world healthcare settings where reporting is incomplete and resource constrained. Third, the results show that failing to account for missing activity data leads to substantial underestimation of hospital GHG emissions and that emissions vary systematically by hospital size and emission source. Together, these contributions provide a scalable and transferable framework for estimating hospital-level GHG emissions that can support targeted decarbonization strategies across healthcare systems in the United States and in other countries where healthcare activity data are incomplete.

2. Methodology

2.1. Data

Hospitals participate in the Environmental Excellence Award program by voluntarily submitting detailed data on their consumption and sustainability-related activities. The program, organized annually by Practice Greenhealth (PGH), collects this information through a comprehensive survey. The data reflect hospital activities during the 2020 period, based either on the calendar year (January 1, 2020 to December 31, 2020) or the fiscal year, depending on the hospital's reporting practices. Fiscal year submissions typically followed either a July–June or October–September cycle (e.g., October 1, 2019 to September 30, 2020).

The PGH dataset included hospitals in the United States (n = 280), Puerto Rico (n = 1), and Canada (n = 2) that voluntarily participated in the Practice Greenhealth Environmental Excellence Award program. The sample consisted primarily of general acute care hospitals (n = 239), along with medical office buildings (n = 14) and specialty hospitals (n = 5). The U.S. hospitals in the dataset represent approximately 3.7% of all U.S. hospitals in 2020, when the total number of hospitals nationwide was 7623 in the same year.

Compared with the national distribution of U.S. hospitals, the study sample includes a higher proportion of medium and large hospitals and a lower proportion of small hospitals (Appendix Figure A1). Nationally, hospitals with more than 500 licensed beds account for approximately 5% of U.S. hospitals, whereas they represent about 20% of hospitals in the PGH dataset. Hospitals with 100–199 licensed beds constitute the second largest group in the study sample, representing 19% of participating hospitals, which is comparable to the national share of 18% in this size category. In contrast, hospitals with fewer than 100 licensed beds represent approximately 17% of the study sample, compared with 54% nationally.

The dataset provides granular information across 15 operational sectors, including hospital characteristics, environmental goals, leadership, recycling, waste, chemicals, anesthetic gases, food services, environmentally preferable purchasing, energy, water, building operations, climate, transportation, and data quality indicators. Quantitative data required for estimating greenhouse gas emissions were available for seven sectors used in this study: energy, anesthetics, food, chemicals, waste, water, and transportation. Reported activity data were provided in variable-specific units appropriate to each measure, such as kilowatthours for electricity consumption and pounds for food purchases.

Overall, the survey included responses to 1003 questions, which were processed into 3296 variables, of which 37% were quantitative and 63% were qualitative. Across all sectors, 38 key variables used to estimate GHG emissions were identified and are hereafter referred to as “source” variables (e.g., the food sector included four source variables measuring the quantities of consumption for beef, pork, poultry, and lamb; see the full list in Appendix Table A1). In the Practice Greenhealth dataset, source variables exhibit heterogeneous patterns of missingness across hospitals, posing challenges for estimating hospital-level emissions from multiple sources (Appendix Table A1). After excluding missing observations for source variables, an additional set of 1192 “predictor” variables with less than 50% missingness was identified as inputs for the imputation models.

Among the selected source and predictor variables, implausible zero values were identified and recoded as missing based on fundamental operational characteristics of hospitals. Zero values were treated as missing when they were not physically or operationally plausible for a functioning facility, such as zero electricity use, zero building area, or zero medical waste generation. In contrast, plausible true zeros were retained for variables that reflect optional or infrequently used items, such as lamb consumption. All variables for which zero values were treated as missing, along with the corresponding operational justification, are summarized in Table A4 in the Appendix. This study also incorporated county-level climate zone information from the Commercial Buildings Energy Consumption Survey (CBECS) to account for geographic variation in building energy use associated with weather and climate (CBECS, 2021).

While the data collected from Practice Greenhealth are primarily self-reported by participating hospitals, the survey incorporates internal mechanisms intended to promote data accuracy. Hospitals are required to submit an affidavit attesting to the accuracy of their responses, and the Environmental Excellence Award program is structured as a supportive, non-punitive initiative. Facilities are neither penalized for high emissions nor publicly ranked in ways that could create reputational risk, which reduces incentives for strategic misreporting and encourages reporting for internal performance tracking and improvement.

Independent external verification of reported activity data was not feasible within the scope of this study, a limitation common to large-scale surveys of institutional operations. Rather than treating this constraint as a barrier to analysis, the study adopts a methodological approach designed for real-world policy and planning contexts characterized by incomplete, heterogeneous, and self-reported data. The machine learning–based imputation framework improves robustness and internal consistency across facilities, enabling comparative analysis, identification of high-impact emission sources, and assessment of systematic differences across hospital types.

Accordingly, the emissions estimates produced in this study are intended to inform understanding of hospital-level emissions patterns, support priority setting and benchmarking, and guide the design of data collection systems and mitigation strategies. Given current limitations in the data, the results are not intended to serve as regulatory compliance inventories or as the basis for enforcement actions. With more standardized reporting and improved data completeness, the approach could support more accurate estimation and prediction of hospital-level GHG emissions across multiple sectors (Yin et al., 2023).

2.2. GBM Imputation of Missing Values

To impute missing values in source variables, this study fitted a gradient boosting machine (GBM) model for each variable, following the methodology described in (Yin et al., 2023). This study used GBM as implemented in the gbm package in R 4.1.2 (Greenwell, 2022), because of its strong predictive performance, flexibility in modeling nonlinear relationships, and its ability to handle missing values in predictor variables through surrogate splits (Friedman, 2001). Surrogate splits route observations with missing predictor values using alternative predictors that approximate the primary split, allowing model fitting without pre-imputing predictors or making additional assumptions about missingness mechanisms.

For each source variable, the analysis conducted an exhaustive grid search in which the GBM model was re-fitted across all 108 predefined combinations of hyperparameters, including the number of trees, maximum interaction depth, minimum number of observations in terminal nodes, learning rate (shrinkage), subsampling fraction, and number of cross-validation folds (Appendix Table A2). The same hyperparameter grid was applied consistently across all source variables. For each outcome, the optimal hyperparameter set was selected as the combination that minimized cross-validated mean squared error using K-fold cross-validation. The predefined hyperparameter grid is summarized in Appendix Table A3, and the selected optimal hyperparameters for each source variable are reported in Appendix Table A3.

Given the large number of candidate predictor variables, this study then implemented a recursive variable selection procedure to reduce model complexity and improve predictive performance. An initial GBM model was fitted using all predictors and the selected hyperparameters, and cross-validated R² was calculated. Predictor variables were ranked by their relative influence, and the least influential 5% were removed. The GBM model was re-fitted on the reduced predictor set, and cross-validated R² was re-evaluated. This process was repeated recursively for 20 iterations until approximately 5% of the original predictor variables remained. From all iterations, the analysis selected the predictor subset that achieved the highest cross-validated R², balancing predictive accuracy and model parsimony. The final GBM model was evaluated using R² on the 30% holdout test dataset and used to predict missing values in the corresponding source variable.

2.3. Estimation of Greenhouse Gas (GHG) Emissions

After predicting missing values in the source variables, the analysis estimated GHG emissions associated with resource consumption, including energy use and anesthetic consumption, across hospitals. Emissions were calculated using sector-specific emission factors from the Greenhouse Gas Protocol, the U.S. Environmental Protection Agency (EPA), and published literature to estimate GHG emissions across seven hospital sectors (GHG Protocol, 2017; US EPA, 2023). The overall method used to estimate GHG emissions can be summarized as the following equation (1):

$$Emissio{n}_{i,j}=Consumptio{n}_i\times E{F}_{i,j}$$ (1)

where $i$ represents source variable $i$, and $j$ indicates the type of GHG, including carbon dioxide (CO₂), methane, nitrous oxide, and other greenhouse gases. The emissions from non-CO₂ greenhouse gases were converted to CO₂ equivalent values using 100-year Global Warming Potential (GWP) published by the Intergovernmental Panel on Climate Change (IPCC) fifth assessment report (IPCC, 2015). The total GHG emissions in each hospital were then estimated by aggregating emissions from each type of consumption.

For GHG emissions from food consumption, this study incorporated three sources of emissions, including annual food related carbon emissions, food related supply chain emissions, and food related carbon emissions due to opportunity cost according to published emission factors (Poore and Nemecek, 2018; Searchinger et al., 2018). As land for agriculture can be used for other carbon sink activities, such as forestation and grassland, this study accounted for the opportunity cost of food consumption that provides policy implications for land use choices (Poore and Nemecek, 2018). To calculate the amount of actual meat for in-bone cuts, this study followed the procedure published by World Resources Institute and the Cool Food Pledge (Equations (2)-(3)) (Waite et al., 2019).

$$M_{m,in-bone}=W_m\times (1-P_{m,boneless})\times FB{S}_m$$ (2)

where $M_{in-bone}$ represents the weight of meat in in-bone cuts, $W_m$ is the total weight of meat purchased in the hospital, $P_{m,boneless}$ is the percent of the meat m that is boneless, and $FB{S}_m$ is a ratio of retail weight to the food supply weight of meat in the U.N. Food and Agriculture Organization food balance sheets, which converts the in-bone cuts to actual amount of meat consumed (FAO, 2022). After calculating the amount of meat in the in-bone cuts, this study estimated the GHG emissions from food using the Equation (3):

$$Emission{s}_m=[W_m\times {(1-P_{m,boneless})}^{\star }FB{S}_m+W_m\times P_{m,boneless}]\times E{F}_m$$ (3)

where $Emission{s}_m$ is the estimates of GHG emission for meat type $m$. The other parameters have the same definition as in Equation (2).

Emission factors are critical in estimating GHG emissions accurately. This study accounted for state-level variations in emission factors for electricity, while spatial uniform emission factors were used to estimate GHG emissions for other source variables. However, these emission factors are the best available knowledge to estimate hospital-level climate impacts.

According to the Greenhouse Gas Protocol, this study further grouped the source variables into three scopes (GHG Protocol, 2017). Scope 1 includes direct emissions from sources owned or controlled by the hospital, such as on-site fuel combustion (e.g., fuel oil, diesel, and natural gas) and direct releases of anesthetic gases during clinical use. Scope 2 includes indirect emissions from purchased energy, including electricity, steam, and other externally supplied energy sources. Scope 3 includes indirect emissions from activities outside the hospital's direct control, such as food procurement, chemical use, and waste management. All emission factors, data years, and source references are summarized in Appendix Data.

For most sectors, scope classification is straightforward, and additional details on scope allocation for each source variable are provided in Appendix Table A3. In contrast, scope classification for the transportation sector depends on vehicle ownership status: emissions from hospital-owned vehicles should be categorized as Scope 1, whereas emissions from leased or contracted vehicles should be categorized as Scope 3. However, the Practice Greenhealth dataset contains substantial missing information on vehicle ownership and leasing status, making it infeasible to allocate transportation emissions between Scope 1 and Scope 3 with reasonable confidence. As a result, transportation-related emissions are reported in aggregate and may correspond to either Scope 1 or Scope 3, depending on ownership. This aggregation does not affect total estimated GHG emissions or the interpretation of sectoral results, as transportation-related emissions associated with hospital operations are incurred regardless of ownership status.

3. Results

3.1. Imputation Accuracy

We assessed the accuracy of the GBM models in predicting missing values in 38 source variables across seven sectors and how accuracy varied by the proportion of values that were missing and had to be imputed (Fig. 1). The source variable with the highest and lowest test R² in each sector is labeled. For example, there are four source variables in the waste sector, including municipal solid waste, regulated medical waste, hazardous waste, and recycling. The test R² was the highest and lowest for municipal solid waste and hazardous waste in the waste sector, respectively. The test R² values were lower for source variables with higher rates of missing data, but there was considerable variability around this trend. For example, in the waste sector, both municipal solid waste and hazardous waste had similar missing rates, approximately 2%. However, the respective test R² values were substantially different, with municipal solid waste (test R² = 0.91) and hazardous waste (test R² = 0.41). This difference may be attributed to the substantially higher variability in hazardous waste between facilities compared to municipal solid waste (Appendix Table A2). In the food sector, the missing rate for poultry was approximately 46%, yet its test R² was 0.92. In contrast, lamb data had a missing rate of 60% with a much lower test R² less than 0.1. Additionally, source variables with a higher frequency of zero values tended to exhibit lower test R² compared to variables with similar missing rates, or coefficients of variation. For instance, the missing rates for lamb (60%), fuel oil (66%), and nitrous oxide (39%), had coefficients of variation below 4, however, their test R² values were 0.09, 0.34, and 0.53, respectively. The zero values in these three source variables were 23%, 7%, and 6% respectively, which were larger than those of any other source variable. The low test R² may be attributed to the selected influencing variables not effectively explaining the zero values. The relative influence allows comparison of the importance of a predictor variable relative to all other predictor variables ²⁵.

Fig. 1.

Model imputation accuracy. The figure displays test R² versus the percent missing for each source variable. Labels are provided for the source variables with the highest and lowest test R² in each sector.

3.2. Resource Consumption by Sector and Source, and by Hospital Size

Energy and resource consumption patterns across hospitals were examined. Fig. 2 shows the observed and imputed estimates for selected source variables in each sector, categorized by the number of staffed beds. Most of these source variables (e.g., electricity, beef, and solid waste) were chosen as they represented the largest contributors to GHG emissions in each sector, while desflurane was selected due to its association with the largest global warming potential compared to other anesthetics (Vollmer et al., 2015).

Fig. 2.

Consumption results for selected source variables in each sector combining both observed and imputed activity data.

Hospitals with a larger number of staffed beds often had a smaller proportion of imputed data (in yellow in Fig. 2A) than those with fewer staffed beds. For example, electricity usage was imputed for 16% of small hospitals (e.g., with 25–49 staffed beds), but no large hospitals (e.g., with over 500 staffed beds) required imputation of electricity usage (Appendix Figure A2). Additionally, 45% of large hospitals had missing data on beef consumption, compared to over 60% of hospitals with less than 50 staffed beds. These results suggest that the availability of data varies depending on the size of the hospital. Larger hospitals may have more resources to document the amount of energy, or resource consumption.

As expected, hospitals with larger staffed-bed capacity consumed more resources and generated greater amounts of waste than smaller hospitals (Fig. 2A). In particular, hospitals in the largest size category (≥500 staffed beds, representing approximately 20% of the study sample) exhibited the highest average levels of energy and resource consumption (Fig. 2B), reflecting their greater volume of procedures and treatments. These hospitals consumed an average of 224,300 million British thermal units (MMBtu) of electricity per year (confidence interval [CI]: 175,690–272,910), more than 21 times the average electricity consumption of hospitals with 6–24 staffed beds. Similarly, average beef consumption, indoor water use, and municipal solid waste generation in hospitals with ≥500 staffed beds exceeded those in the 6–24 staffed-bed category by factors of 28, 37, and 36, respectively.

By aggregating the energy consumption per square foot from electricity and natural gas (PGH, 2020), the average energy use intensity (EUI) for each hospital was 202.9 (197.0, 208.8) thousand British Thermal Units per square foot (KBtu/sqft). The EUI was calculated based on site energy consumption, which does not include energy consumed in transmission, delivery and production losses (Energy Star, 2023a). Specifically, hospitals with 400–499 staffed beds (but not hospitals with 500+ beds) had the largest average EUI of 263.4 (200.3, 326.5) KBtu/sqft (Appendix Figure A3).

Despite consuming fewer total resources, smaller hospitals exhibited higher resource use and GHG emissions per staffed bed compared to larger hospitals. As shown in Fig. 2C, hospitals with fewer than 25 staffed beds had higher per-bed consumption of electricity, beef, indoor water, gasoline for transportation, desflurane, and general-purpose cleaners than hospitals in larger size categories. This pattern likely reflects a combination of structural and operational factors, including limited economies of scale, older or less energy-efficient buildings, less advanced building management and monitoring systems, and data representation. The hospital-level estimates allow individual facilities to examine their own emissions profiles and assess how local infrastructure, service mix, and operational practices contribute to observed emissions. Accordingly, opportunities to reduce energy and resource intensity in smaller hospitals are likely to depend on context-specific strategies tailored to facility characteristics and data capabilities.

In Fig. 2, “activity” refers to the amount of consumption, or waste generated in each sector. Fig. 2A shows the total annual activity-related consumption, or waste generated, categorized by staffed beds. Fig. 2B and C shows the activity per hospital and per staffed bed, respectively. Error bars represent the 95% confidence interval for the mean activity per hospital or per staffed bed, calculated across all hospitals. The colors in Fig. 2B and C corresponds to each source variable. The units for each source variable are labeled in parentheses. The data included total water consumption and irrigation water consumption; indoor water consumption was calculated by subtracting irrigation water from total water consumption.

3.3. Estimate of GHG Emissions

Based on the imputed results, hospitals in the study sample (N = 283) were estimated to emit 9.3 megatonnes (Mt) carbon dioxide (CO₂) equivalent emissions across the seven sectors examined, which represented 1.9% of the previously reported GHG emissions from the US healthcare industry in 2018, using a top-down approach based on Environmentally Extended Input-Output (EEIO) models (Eckelman et al., 2020). The GHG emissions from scope 2 were 6.2 megatonnes (Mt) CO₂ equivalent, followed by scope 1 (2.0 megatonnes (Mt) CO₂ equivalent) and scope 3 (1.0 megatonnes (Mt) CO₂ equivalent). When missing activity data are treated as zero, hospital-level GHG emissions are systematically underestimated because emissions from unreported activities are excluded by construction. Accounting for missing activity data through imputation restores these omitted contributions. As shown in Figs. A4 and A5 in the Appendix, emissions associated with imputed activity data account for approximately 37% of total estimated GHG emissions in the study facilities.

Compared to estimates derived from the EEIO modelling (Eckelman et al., 2020), emissions estimated from scope 3 sources represent a considerably smaller share of the total emissions. This is because there were few scope 3 activities data included in the PGH dataset. Scope 3 emissions encompass diverse activities across healthcare delivery, and emission factors attributable to these activities are not well documented. For example, scope 3 emissions include pharmaceuticals, however assigning an accurate carbon footprint is challenging due to the lack of data available on the materials and methods used in manufacturing (Parvatker et al., 2019).

Across the seven sectors analyzed, energy, food, and waste were the dominant contributors, together accounting for approximately 99% of total estimated GHG emissions. The distribution of emissions in each sector is shown in Appendix Figure A6. Energy use was by far the largest contributor, followed by food, waste, anesthetics, transportation, water and chemicals. The distribution of hospital GHG emissions by selected sources in each sector is shown in Fig. A6. Electricity consumption was the largest contributor to the energy sector (2.4 megatonnes (Mt) CO₂ equivalent); beef consumption was the largest source of GHG emissions in the food sector (0.6 megatonnes (Mt) CO₂ equivalent). Desflurane (28,000 tonnes (t) CO₂ equivalent), in the anesthetics sector, contributed more than gasoline-related emissions (26,000 tonnes (t) CO₂ equivalent) in the transportation sector. Although desflurane has the most potent GHG effect (Vollmer et al., 2015), total GHG emissions from nitrous oxide were over 100,000 tonnes (t) CO₂ equivalent in this study, which exceeded those from desflurane by a factor of 2.6 (Appendix Figure A12). Nitrous oxide constituted nearly 70% of total GHG emissions associated with anesthetics in this study, a higher share than reported anesthetic emissions in the UK, where nitrous oxide accounts for approximately 30% (Chakera et al., 2022). This difference may reflect variation in anesthetic practices, clinical protocols, and medical gas management systems between the two countries. In addition, the relative contribution of nitrous oxide may reflect changing anesthetic use patterns in U.S. hospitals, particularly the reduction or phase-out of desflurane due to its high global warming potential, which increases the proportional contribution of nitrous oxide among remaining anesthetic agents. There is also notable uncertainty in the estimated nitrous oxide emissions, as indicated by a moderate test R² value (0.53), which may influence its estimated share within total anesthetic-related emissions. Detailed source-specific GHG emissions by sector are presented in Appendix Figures A7-A13.

The average GHG emissions per hospital from the seven sectors were on the order of 33,000 tonnes CO₂ equivalent (CI: 29,460, 36,333)]. Larger hospitals generated more GHG emissions than smaller hospitals (Appendix Figures A14-A15). For instance, hospitals with less than 25 staffed beds emitted on the order of 15,000 tonnes (t) CO₂ equivalent per hospital (CI: 13,744, 15,355), while those with over 500 staffed beds emitted on the order of 74,000 tonnes (t) CO₂ equivalent per hospital (CI: 60,303, 87,722).

Emissions were also normalized by hospital size using the number of staffed beds (Fig. 3). Among the selected sources in Fig. 3, GHG emission per staffed bed from electricity consumption was 38 tonnes (t) CO₂ equivalent (34, 42), followed by beef (8.5 tonnes (t) CO₂ equivalent (8.0, 9.0)) and municipal solid waste (2.0 tonnes (t) CO₂ equivalent (1.9, 2.1)).

Fig. 3.

GHG emissions per staffed bed by selected source variables in each sector. The colors represent sector, or source in corresponding sector.

While the total carbon footprint was positively correlated with the number of staffed beds, GHG emissions per bed from selected sources decreased with size for all except municipal solid waste (Fig. 4), again suggesting that larger facilities are more efficient in managing their energy and resources consumption. Specifically, smaller healthcare facilities with fewer than 25 staffed beds emitted 903 tonnes (t) CO₂ equivalent (567, 1239) per staffed bed, which was over eight times higher than that of larger hospitals with over 500 staffed beds.

Fig. 4.

Correlations between GHG emissions per staffed bed and the number of staffed beds in different source variables. The regression functions indicate the relationship between natural log transformed GHG emissions per staffed bed (y) and the number of staffed beds (x).

While the number of staffed beds exhibited a significant correlation with GHG emissions per staffed bed across most sources, the regressions in Fig. 4 show that the R² values were small for most sources. This implies that substantial unexplained variation that cannot be captured by the number of staffed beds.

Correlations between GHG emissions per staffed bed and the number of staffed beds in each sector are shown in Appendix Figures A16-A22. Of all sources, only regulated medical waste GHG emissions per bed increased with the number staffed beds (Appendix Figure A18). It is possible that the type of patients and procedures in larger hospitals result in generation of more regulated medical waste per patient. In general, there was substantial variation in GHG emissions per staffed bed across almost all source variables.

While staffed beds indicate a hospital's current operational capacity, licensed beds represent its maximum capacity. As a sensitivity analysis, GHG emissions were normalized by the number of licensed beds, yielding similar correlation patterns, which suggest that larger hospitals generally had lower GHG emissions per licensed bed (Appendix Figure A23).

In addition to hospital size, this study compares emissions between academic and non-academic hospitals. Average GHG emissions per hospital in academic hospitals were 96% higher than those in non-academic facilities (Appendix Figure A24). Academic hospitals had 39% lower emissions per staffed bed compared to non-academic hospitals (Appendix Figure A25). However, academic facilities were generally much larger than non-academic facilities (with over twice the number staffed beds). Differences were markedly attenuated by adjusting for number of staff beds and were no longer statistically significant (Appendix Table A5).

3.4. Factors Influencing Resource Consumption and GHG Emissions

Factors contributing to heterogeneity in resource consumption and waste generation were examined for selected sources. Building area, amount of municipal solid waste, number of patient days, and spending on desflurane were the leading factors associated with electricity use, beef consumption, municipal solid waste generation, and desflurane use, respectively (Fig. 5).

Fig. 5.

The relative contribution of influencing factors for selected sources: (A) Electricity, (B) Beef, (C) Solid Waste, and (D) Desflurane, are presenting in each panel. Relative influence (left): A higher value indicates greater importance of a variable in predicting the source variable. The colors in each panel represent the influencing factors, which have been re-grouped into hospital characteristics and corresponding sectors in the stacked bar plot (right). Hospital characteristics: Structural and operational attributes of facilities.

The stacked bar plots in Fig. 5 (on the right) depict the aggregated effects of influencing factors that were used to impute sources, categorized according to the 15 sectors in the PGH survey. Energy-related factors, such as building area and natural gas use, collectively have the greatest influence on electricity consumption. Relative influence in the figure quantifies each predictor's contribution to reducing prediction error across all trees in the ensemble. Higher relative influence values indicate that a variable plays a larger role in improving model predictions, whereas lower values indicate a more limited contribution to explaining variation in the outcome. Hospital characteristics denote structural and operational attributes of facilities, such as building area, number of beds, operating rooms, patient volume, and staffing levels, rather than activity-based consumption variables.

According to the partial dependence results, larger building areas were associated with higher electricity consumption, while greater natural gas use was associated with lower electricity consumption in healthcare facilities (Appendix Figure A26). This pattern likely reflects facilities using natural gas for space heating and water heating, which reduces the need for electricity for these end uses. Factors related to hospital characteristics, waste, and food categories influenced beef consumption. Hospitals with more municipal solid waste also had higher consumption of beef, and a higher spend on food and beverages was also correlated with higher beef consumption (Appendix Figure A27). Similarly, hospitals with more staffed beds, or licensed beds tended to use more beef in food services. Municipal solid waste was also highly impacted by hospital characteristic such as patient days and licensed beds (Appendix Figure A28). Hospitals with higher adjusted patient days also had higher numbers of gasoline vehicles (Appendix Figure A29). Desflurane consumption was primarily influenced by greening the operating rooms (greening OR) related factors (Fig. 5), including amount spent on desflurane and the annual number of operating room procedures (Appendix Figure A30). Additional partial dependence plots are included for total water and general purpose cleaners in Appendix Figures A31 and Figure A32.

4. Discussion

Monitoring of GHG emissions at the hospital level is an important tool for a hospital to be able to intervene to reduce its carbon footprint. By aggregating facility-level operational data collected through Practice Greenhealth, this study demonstrates the feasibility of a bottom-up approach to hospital-level GHG emissions accounting as a complement to national and subnational top-down estimates. Using a GBM–based framework to address missing data (Yin et al., 2023), this study estimated GHG emissions across seven operational sectors for 283 hospitals, representing approximately 3.7% of U.S. healthcare facilities (HIFLD, 2020). Although the study sample overrepresents larger hospitals, the estimated emissions account for only 1.9% of prior national healthcare emissions estimates (Eckelman et al., 2020), underscoring that the results represent a conservative subset of total healthcare emissions.

4.1. Hospital Scope 3 Emission Sources

The seven sectors constitute a substantial portion of total GHG emissions, but do not account for emissions from various other sources due in large part to limited data on scope 3 sources, such as pharmaceuticals, single-use medical devices, and other supply-chain activities (Eckelman et al., 2020). Many hospitals currently lack the data collection systems needed to report these activities at the facility level. Thus, the estimates of total GHG emissions presented here likely underestimate the full climate impact of the hospitals studied, particularly for Scope 3 sources. With improved and more standardized hospital-level collection of supply chain–related activity data, the analytical framework developed in this study could be extended to more accurately estimate and predict GHG emissions from major scope 3 contributors. Enhancing data availability through standardized reporting procedures would substantially reduce uncertainty in healthcare supply chain emissions estimates.

Across the sectors examined, the food sector emerged as a substantial contributor to hospital GHG emissions, particularly when land-use–related opportunity costs are considered. Opportunity cost reflects the potential carbon sequestration foregone due to agricultural land use and provides insight into the broader climate implications of food procurement choices, especially for emissions-intensive products such as beef and lamb. For example, when a hospital replaces imported beef-based meals with poultry alternatives, supply-chain emissions alone may suggest a notable reduction in GHG emissions. However, accounting for opportunity cost reveals substantially larger mitigation potential by freeing land that could otherwise support carbon sequestration. Ignoring opportunity cost in this context may therefore lead hospitals to underestimate the climate benefits of dietary shifts and prioritize less effective mitigation strategies.

To assess the influence of this methodological choice, a sensitivity analysis was conducted in which food-related emissions were estimated both with and without opportunity cost. When opportunity cost was excluded, the food sector contributed an estimated 144,498 tonnes (t) CO₂ equivalent across the 283 hospitals, approximately 80% lower than estimates that include opportunity cost (757,825 tonnes (t) CO₂ equivalent). This comparison illustrates that opportunity cost substantially affects the estimates of food-sector emissions. When opportunity cost is included, the food sector accounts for approximately 8% of total estimated hospital emissions (1.6% when excluded), a contribution comparable in magnitude to food-related emissions reported in national healthcare assessments in the UK (approximately 6%) (Sadler et al., 2025) and Germany (4–7%) (Keil, 2023), and smaller than estimates reported for Switzerland (17%) (Keller et al., 2021). Differences across studies likely arise from variation in system boundaries, national or regional context, clinical practices, analytical scope, and underlying data availability. Additionally, the seven sectors analyzed capture a substantial share of hospital emissions but exclude several important Scope 3 sources, such as pharmaceuticals, single-use medical devices, and other upstream supply-chain activities, due to limited activity data in the PGH survey. Consequently, the total GHG emissions estimated here likely underestimate the full climate impact of the hospitals studied, and the relative contributions of the food sector and other included sectors should be interpreted with caution and may not be directly comparable to studies with different system boundaries. However, with improved and more standardized collection of supply chain–related activity data, the analytical framework developed in this study could be extended to more accurately estimate and predict GHG emissions from major Scope 3 contributors, thereby reducing uncertainty and supporting more effective hospital decarbonization strategies.

4.2. Predictor Variables

An important finding of this analysis is that readily available administrative and structural characteristics of hospitals, such as building area, number of operating rooms, staffed and licensed beds, physician counts, and patient occupancy, are strong predictors of resource consumption and associated GHG emissions. Because these variables are routinely recorded in administrative datasets, they offer practical value for hospital-level emissions accounting and benchmarking, particularly when detailed activity data are incomplete. This finding demonstrates that administrative data can support systematic estimation of hospital emissions, even though such variables are not direct levers for mitigation. For example, identifying an association between operating room capacity and emissions does not imply that reducing operating rooms is a feasible or appropriate intervention. When combined with partial activity data, administrative variables can also improve prediction of emissions from sources that are not directly reported, supporting more complete emissions inventories. Due to the representativeness of the current survey sample and the high rate of missing data, the models estimated here are not intended for direct application beyond the study sample. However, the modeling framework applied in this study can be used to develop predictive models for estimating GHG emissions in other hospitals facing similar data limitations. With improved data quality, more standardized reporting practices, multi-year observations, and broader participation in the PGH survey or other data collection systems, the GBM framework could be used to re-estimate models with enhanced predictive performance and generalizability across US hospitals. Continued improvements in data infrastructure would enable this approach to support more accurate, transferable, and actionable hospital-level decarbonization strategies.

4.3. Hospital Size

This study demonstrated the utility of the modeling approach to show that larger hospitals had, not surprisingly, larger GHG emissions but that the carbon footprint per staffed, or licensed bed was generally reduced as the size of the hospital (number of beds) increased across most sources studied, suggesting that larger hospitals are more efficient in use of resources that cause GHG emissions. An exception to this pattern was observed for regulated medical waste, for which consumption and emissions per staffed bed increased with hospital size. In addition, lower emissions per bed in academic medical centers were largely explained by the size of the hospital. These findings highlight areas for further investigation and demonstrate that the modeling framework applied in this study can be used to test hypotheses related to hospital resource use and emissions.

4.4. Prediction of Emissions

The results of the imputation analysis indicate that the precision of imputation tends to be higher for source variables with a lower missing rate and less variability. Some source variables, such as poultry, municipal solid waste, pork, water consumption, and electricity, demonstrated good prediction precision of the GBM models with a test R² larger than 0.8. These findings suggest that, with more training data, it would be possible to better predict resource consumption and estimate GHG emissions in hospitals that lack complete data.

While the GBM framework applied in this study provides a promising approach for predicting hospital carbon footprints, the models are relatively data intensive and typically required more than 20 influencing variables to achieve reliable predictions. This reflects not only the complexity of hospital operations, but also the high rate of missing values in many individual predictors, which necessitates the use of a broader set of variables to maintain predictive accuracy. Future efforts to develop models that rely on fewer input variables, supported by improved data completeness and standardization, would be valuable for advancing hospital-level GHG emissions estimation across the healthcare sector.

4.5. Limitations

Although this study provides a first large-scale, bottom-up assessment of hospital-level GHG emissions using a machine learning approach to address common missing activity data in hospitals, several limitations should be noted. Model performance varied across emission sources. While the GBM models achieved strong predictive performance for 10 source variables, with test R² values greater than 0.7, 16 source variables had test R² values below 0.4, indicating greater uncertainty in the corresponding imputed estimates. This uncertainty largely reflects high rates of missing activity data in both source and predictor variables, highlighting the importance of improved and more standardized hospital-level data collection. In addition, the study sample disproportionately represents larger hospitals, with 72% of facilities having more than 100 staffed beds compared with 42% nationally in 2020 (HIFLD, 2020). Consequently, the imputation results may be less representative of smaller hospitals, and caution is advised when generalizing the findings to the broader hospital population.

The unprecedented impacts and disruptions to healthcare systems from COVID-19 undoubtedly influenced the GHG emissions reported in 2020. The surge of infected cases stretched the capacity of healthcare facilities across the US (Mareiniss, 2020), where hospitals were forced to postpone, or cancel elective operations and non-urgent procedures to avoid the spread of the infection (Brown et al., 2021). To assess the influence of COVID-19 related variables on the prediction model, this study explored the total volume of desflurane by removing these variables in the GBM model. The results showed a slight increase (1.4%) (Appendix Figure A34). The sensitivity test illustrates potential impacts of the pandemic on resource consumption in healthcare facilities.

Energy consumption may also have been affected by the disruption of the pandemic. The average EUI calculated in this study was 202.9 (197.0, 208.8) KBtu per square foot (KBtu/sqft), which was 5% higher than the national average (193.3 KBtu/sqft) for all inpatient healthcare buildings in the US, as reported in the Commercial Buildings Energy Consumption Survey (CBECS) in 2018 (CBECS, 2022) (Appendix Figure A3). In comparison to the 2015 Energy Star report, which was based on an industry survey conducted by the American Society for Healthcare Engineering (ASHE), the EUI was 13% lower than the reported value of 234.3 KBtu/sqft (Energy Star, 2023b). The differences in EUI can be attributed to the influence of the pandemic. For instance, where hospitals experienced severe impacts due to a surge in COVID-19 cases, nearly all building spaces, except operating rooms, were operating at full capacity. Consequently, energy consumption during this time likely differed significantly from a typical year without such disruptions. The hospital sample included in this study is much smaller than these two national surveys, which could introduce bias into the average EUI estimates. Lastly, district heat and fuel oil were also excluded from the EUI calculation in this study, whereas both energy sources are included in EUI estimates reported by CBECS and Energy Star. In summary, the variation in EUI values can be attributed to the complex interaction of pandemic-related disruptions, sample size differences, survey timing variations, and differences in calculation methods.

The influence of the pandemic might also explain why municipal solid waste per staffed bed demonstrated less variation across healthcare facilities (Fig. 4). During the pandemic, hospitals were primarily focused on treating COVID-19 patients with similar conditions. Thus, healthcare facilities may generate comparable amounts of municipal solid waste. Additionally, food waste typically constitutes 20%–30% of the total waste in healthcare facilities (Carino et al., 2020), however food services in many healthcare systems were closed during the pandemic (see Appendix Table A6), which may have contributed to reducing variations in municipal solid waste. Therefore, the patterns of resource consumption and waste generation might have deviated from the typical scenario without the pandemic, and the estimated GHG emissions may not reflect typical operations.

5. Policy and operational implications

The findings of this study have several implications for healthcare decarbonization efforts at the hospital, healthcare-system, and policy levels. Hospital-level GHG emissions estimates can help administrators identify high-impact sectors, such as energy use, food services, and waste management, and prioritize targeted mitigation strategies, including energy efficiency investments, low-carbon procurement, and operational changes that reduce resource intensity. By highlighting systematic differences in emissions intensity and data completeness by hospital size, the results underscore the importance of tailoring decarbonization strategies to institutional capacity. Smaller hospitals often face constraints related to staffing, technical expertise, and financial resources, which can limit their ability to conduct detailed emissions inventories or implement capital-intensive mitigation measures. For these facilities, standardized GHG accounting guidance, simplified reporting tools, peer-learning networks and access to shared data infrastructure may be particularly valuable for improving emissions measurement and identifying priority actions.

At the health-system and policy levels, the analysis highlights the need for standardized data collection systems and sector-wide reporting frameworks to improve the consistency, comparability, and usability of hospital activity data. While smaller hospitals tend to face greater challenges due to limited staffing and technical capacity, incomplete reporting is not unique to small facilities; large hospitals also exhibit missing data across key sectors. Centralized hospital financial systems, integrated activity recording platforms, and streamlined, standardized approaches to tracking resource use and operational activities could substantially reduce reporting burdens across hospitals of all sizes and improve the accuracy of GHG emissions accounting. Without coordinated, system-level data standardization and streamlined data collection, reporting burdens may continue to limit large-scale facility- and sector-level GHG emissions accounting, thereby hindering hospitals’ ability to reduce their carbon footprint.

The findings also suggest that governance and operational models play an important role in shaping the adoption of low-carbon practices in healthcare. Voluntary reporting programs, such as Environmental Excellence Award program administered by Practice Greenhealth, can encourage early participation, learning, and capacity building across hospitals. However, achieving broader and more consistent emissions reductions is likely to require policy frameworks that combine clear standards with incentives and flexibility. In particular, market-based approaches, such as cap-and-trade systems or carbon pricing mechanisms, offer cost-effective pathways for emissions reduction by allowing hospitals to select mitigation strategies based on their marginal abatement costs (Goulder and Schein, 2013; Stavins, 1998, 2003). While these policies may raise concerns about potential cost pass-through to healthcare services, revenue recycling mechanisms can be designed to mitigate distributional impacts, for example by subsidizing patient costs or supporting investments that reduce long-term operating expenses (Stavins, 2003). These approaches can accommodate substantial heterogeneity across facilities while preserving overall efficiency in achieving emissions reduction goals. Together, voluntary programs, standardized reporting, and market-based policy instruments can help translate hospital-level emissions accounting into coordinated and efficient mitigation action across the healthcare sector.

6. Conclusion

In conclusion, this study applied a machine learning–based framework to estimate hospital-level GHG emissions from incomplete self-reported data, providing a bottom-up assessment across seven key operational sectors for 283 hospitals. The estimated emissions of 9.3 megatonnes (Mt) CO₂ equivalent represent approximately 1.9% of previously reported U.S. healthcare emissions. The analysis demonstrates that machine learning–based imputation can effectively address missing data challenges and enable systematic emissions estimation at scale, while identifying associations between key factors and hospitals carbon footprint. While reducing emissions in larger hospitals remains critical, the lower resource efficiency observed in smaller hospitals suggests that they also represent important opportunities for emissions reduction, particularly through improvements in resource efficiency and data reporting practices. At the same time, limitations in data availability highlight the need for more standardized and comprehensive data collection to improve the accuracy of hospital-level emissions estimates. Overall, this study provides a scalable and transferable framework for hospital-level GHG emissions accounting. Continued improvements in data infrastructure and the implementation of evidence-based strategies will be essential to support effective decarbonization efforts while maintaining the quality of healthcare services.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.cesys.2026.100461.

Data sharing

The PGH survey data used in this study were not publicly available due to the data use policy implemented to ensure the confidentiality of sensitive information in the dataset.

Data availability

The data that has been used is confidential.