Predicting the Climate Impact of Healthcare Facilities Using Gradient Boosting Machines

Abstract

Health care accounts for 9–10% of greenhouse gas (GHG) emissions in the United States. Strategies for monitoring these emissions at the hospital level are needed to decarbonize the sector. However, data collection to estimate emissions is challenging, especially for smaller hospitals. We explored the potential of gradient boosting machines (GBM) to impute missing data on resource consumption in the 2020 survey of a consortium of 283 hospitals participating in Practice Greenhealth. GBM imputed missing values for selected variables in order to predict electricity use and beef consumption (R²=0.82) and anesthetic gas desflurane use (R²=0.51), using administrative data readily available for most hospitals. After imputing missing consumption data, estimated GHG emissions associated with these three examples totaled over 3 million metric tons of CO₂ equivalent emissions (MTCO₂e). Specifically, electricity consumption had the largest total carbon footprint (2.4 MTCO₂e), followed by beef (0.6 million MTCO₂e) and desflurane consumption (0.03 million MTCO₂e) across the 283 hospitals. The approach should be applicable to other sources of hospital GHGs in order to estimate total emissions of individual hospitals and to refine survey questions to help develop better intervention strategies.

1. Introduction

Healthcare facilities provide complex medical services that consume significant amounts of energy and resources. The US healthcare market is one of the world’s largest, accounting for over 17% of the nation’s gross domestic product and 9% - 10% of national greenhouse gas (GHG) emissions (Hartman et al., 2020; J. D. Sherman et al., 2020).

To assess ways to reduce their climate impact, hospitals require detailed audits on their current resource consumption and associated carbon footprint. Healthcare GHG emissions are often categorized into three scopes: direct emissions from energy consumption (scope 1), emissions from purchased electricity and energy from external sources (scope 2), and indirect emissions from supply chains, waste disposal, transportation, and construction (scope 3). Accounting for emissions from each of these scopes requires data from a wide range of sectors within a healthcare facility and outsourced services, including energy usage, pharmaceuticals consumption, medical devices, waste management, catering services, transportation, and construction (Salas et al., 2020; Singh et al., 2022; Wu, 2019).

Practice Greenhealth (PGH) is a non-profit organization that collaborates with over 1400 U.S. hospitals and healthcare facilities to encourage sustainable practices towards the goal of reducing GHG emissions using emerging evidence of best practices, many of which can result in substantial savings (HCWH & ARUP, 2019). To track the environmental impact of its partner facilities, PGH provides guidance for reporting hospital environmental performance through an annual Environmental Excellence Awards survey. The survey comprises an inventory of activities that influence the carbon footprint, offering valuable insights into GHG emissions accounting for healthcare facilities. However, many hospitals struggle with collecting, analyzing, and reporting these data (Eckelman et al., 2020; Tennison et al., 2021). For instance, healthcare facilities often face resource limitations in collecting data on energy consumption, anesthetic gas usage, food consumption, and healthcare-related waste generation. The absence of such data can hinder the ability to implement effective strategies for reducing hospital climate impact.

Addressing the challenge of missing consumption data is essential for accurately accounting for the carbon emissions of PGH hospitals that participate in the survey. There is limited research that explores methods to impute missing consumption data at the facility level. To tackle this challenge, we have developed a machine learning approach that utilizes the unique dataset from the PGH survey. This approach provides a useful framework for imputing missing data and estimating the carbon footprint of hospitals participating in surveys like PGH’s. We selected the Gradient Boosting Machine (GBM) approach due to its ability to produce accurate predictions, even in the face of missing covariate data (Azizi & Hu, 2020; Natekin & Knoll, 2013). To our knowledge, this study is the first to use GBM to predict missing data on resource consumption contributing to GHG emissions in hospital facilities.

To demonstrate our approach to the missing data problem, in 283 healthcare facilities responding to the 2020 PGH survey awe applied GBM to predict three selected sources of GHG emissions: usage of the anesthetic gas desflurane, a potent GHG (Scope 1), electricity consumption (Scope 2), and purchased beef (Scope 3). These sources were chosen because they are already known to be relatively significant contributors to greenhouse gas emissions in the anesthetics, energy, and food sectors of healthcare facilities (McGain et al., 2020; Poore & Nemecek, 2018; J. Sherman et al., 2012; Tennison et al., 2021). They also serve as representative sources in three emission scopes. We developed predictive models by training the GBM model on PGH data with observed usage from healthcare facilities and using the model to impute usage in facilities with missing data. Then, using both observed and imputed data, we estimated facility-level GHG emissions and across the consortium from the three selected sources. Our approach provides practical solutions for consortia of healthcare facilities to apply machine learning tools to estimate facility-level GHG emissions despite missing data, important information needed to design effective decarbonization strategies.

2. Materials and Methods

2.1. Data and data preprocessing

The 2020 PGH Environmental Excellence Awards survey was submitted by 283 hospitals in the US in 2021; 80% were medium to large hospitals (>=100 licensed beds). It is worth noting that the data collection periods varied among hospitals. For the 2020 data, approximately 64% of the hospitals collected their data based on a calendar year (1/1/2020 to 12/31/2020), while around 34% collected the data on a fiscal year basis, which followed either a July-June or October-September cycle (e.g., 10/1/2019 to 9/30/2020). The PGH survey covered 15 domains, including: hospital characteristics, environmental goals, sustainable leadership, recycling, waste management, chemicals consumption, operating room practices, food services, environmentally preferable purchasing, energy consumption, water management, building design and maintenance, climate impacts, transportation, and self-reporting of the accuracy of the data. The survey contained 1003 questions which were processed into 3296 variables available for analysis.

A major challenge in quantifying GHG emissions using these survey data is the substantial proportion of missing information, largely due to hospital-level challenges in acquiring detailed documentation on energy consumption and procurement. As shown in Appendix Figure A1, there was a wide range in the missingness of each of the 3296 variables in the survey. There were 294 variables (9% of total) missing less than 10% of values across the 283 hospitals, and 378 variables (11% of total) missing greater than 90%.

To develop the imputation model, we focused on the three selected sources of GHG emissions (from Scopes 1, 2, and 3) that also represented a range of missingness: electricity was missing 5% of values across hospitals, beef 46%, and desflurane 56%.

Data were preprocessed separately for each source of GHG to be imputed (desflurane, electricity, beef), henceforth called the “outcome variables”, according to the following steps (Appendix Figure A2). First, we removed all the secondary variables that were calculated based on primary variables to avoid collinearity. Second, we subset the data to include hospitals with non-missing information on the outcome variables. Because each of the selected outcome variables was right skewed, we log-transformed them prior to modeling to reduce the influence of extremely large values. Third, we manually curated the set of potential predictor variables, acknowledging the limitations of the survey data by: (a) excluding non-outcome predictor variables with <50% missing values and (b) amongst this subset of potential predictor variables identifying and replacing potentially false zero values with missing values (e.g., hospitals reporting 0 outpatient visits or 0 waste generated had the 0 values replaced with missing). The final analysis datasets contained: 873 predictor variables for 125 hospitals for desflurane, 867 predictor variables for 269 hospitals for electricity, and 748 predictor variables for 154 hospitals for beef. Finally, we randomly partitioned the final analysis dataset for each outcome variable into 70% training and 30% test datasets. We used the training dataset to develop the GBM models and evaluated their performance using the test dataset.

2.2. Gradient boosting machine models

GBM is a machine learning method based on a sequence of decision trees, developed from learning theory and using the gradient boost algorithm (Freund, 1995; Friedman, 2001). GBM builds trees consecutively to reduce the residual errors of previous trees, correcting previous mistakes (Y. Zhang & Haghani, 2015). The algorithm in the GBM framework works by partitioning the input variables to minimize the loss function, thereby achieving the most homogeneous responses to the predictors (Touzani et al., 2018; Z. Zhang et al., 2019). GBM has been used widely in real-world applications, including disease prediction, diagnosis, treatment, and healthcare management (Cho et al., 2019; Luo et al., 2019; Spann et al., 2020).

Compared to other machine learning tools, GBM has four major advantages. First, GBM has excellent prediction, typically outperforming regression and other machine learning models (Touzani et al., 2018) due to the sequential construction of trees focused on correcting previous errors (Friedman, 2001). When compared to other machine learning models like Random Forest and support vector machines, GBM consistently demonstrates higher prediction accuracy (Golden et al., 2019; Statnikov et al., 2008; Yoon, 2021). Second, GBMs also demonstrate a high-level of efficiency even when fitting with thousands of input variables (Bühlmann & Hothorn, 2007). Their efficiency is derived from the iterative nature of GBMs, where each iteration builds upon the knowledge of the previous learner (Konstantinov & Utkin, 2021). Given the presence of over 3000 variables in the PGH survey data, the GBM model is well-suited for predicting missing values in the dataset. Third, GBMs can flexibly handle different types of data, including continuous, binary, and categorical variables, and can represent nonlinear relationships and interactions through sequential trees (Hastie et al., 2009; Natekin & Knoll, 2013). As a result, GBM models require minimal data preprocessing when dealing with the diverse data types present in the PGH dataset. Fourth, GBMs can handle missing data in explanatory variables using a surrogate split approach (Coffman et al., 2020; Friedman, 2001). For each split in the tree, GBMs consider other variables that might serve as good substitutes for the one with missing data. These surrogate variables are selected based on their ability to mimic the original split. This feature is particularly relevant here, as missing data was prevalent across outcome and explanatory variables. We implemented GBM using the “gbm” package (Brandon Greenwell, 2022) in R version 4.1.2.

2.3. GBM hyperparameter tuning

The performance of a GBM model depends on the choice of hyperparameters, which include: number of trees, maximum tree depth, learning rate (shrinkage), minimum number of observations in terminal nodes, number of cross-validation (CV) folds, and bagging fraction (Natekin & Knoll, 2013). The number of trees is the total number of decision trees used for the final prediction, with too many trees leading to overfitting. The maximum tree depth is the maximum number of splits allowed within each tree, with a greater number of splits allowing for additional flexibility (i.e., nonlinearity, interactions) and potentially reducing the number of trees required. The minimum number of observations in the terminal nodes of each tree impacts the number of splits, and when set too low results in overfitting. The learning rate (shrinkage) controls the contribution of each tree to the final prediction, with a low learning rate avoiding overfitting at the cost of slower convergence. In GBM, CV is used to estimate the out-of-sample performance of the model and avoid overfitting. The bagging fraction is the proportion of the training sample selected in a random subsample to improve model stability and reduce overfitting.

To select optimal hyperparameters we used a grid-search by selecting several values for each hyperparameter (Appendix Figure A3) and re-running GBM on each of the 108 combinations of hyperparameters (Appendix Table A1) to identify the combination minimizing the mean squared error (MSE) criterion.

2.4. Variable selection

Variable selection, also known for feature selection, is an important step in the application in machine learning models (Pfeifer et al., 2022), particularly when dealing with datasets containing a large number of variables (Guyon & Elisseeff, 2003). The process of selecting variables enhances the ease of visualizing and interpreting the results, and more importantly reducing training time and improving prediction performance. In this study, the final dataset for each outcome variable has hundreds of input variables. To evaluate whether a smaller set of predictor variables could produce GBM models with good predictive ability, we implemented the following 4-step recursive variable selection strategy(X. Chen & Jeong, 2007), separately for each outcome:

1. Fit the initial GBM with all predictor variables, using optimal parameters obtained from hyperparameter tuning. Then, we used the trained model to predict outcome variables for the test dataset, and calculated training R², cross validation (CV) R², and test R².

2. Reduction in predictor variables. We ranked the predictor variables by their relative influence in the initial model and removed the bottom 5% variables in the training dataset.

3. Recursively fit GBM models. We repeated the second step with the new training dataset to calculate new R² values and continue the variable reduction process recursively until 5% variables were left. In total, we ran 20 iterations of GBM models and predictions across all selected variables.

4. Optimal variables selection. We selected the set of predictor variables that gives the highest CV R² from all GBM models. The selected variables were then used as inputs in the final GBM model.

2.5. Final model evaluation and imputation of missing values

After building the final GBMs using the 70% training data, we assessed their performance by using the 30% test dataset to calculate the test R². To further evaluate the models, we examined the relative influence plots and partial dependence plots. Relative influence demonstrated the importance of each predictor variable in the model. The relative influence of a predictor variable represents the percent increase in the MSE when the predictor variable is randomly permuted. The higher the relative influence of a variable, the more important it is in predicting the outcome variable. Partial dependence plots illustrate the relationship between a predictor variable and the outcome variable while controlling for the other predictor variables. These plots allowed us to analyze the individual impact of each predictor variable on the outcome variable. The final GBMs were then used to predict missing values in the outcome variables, so that all 283 hospitals had either an observed or imputed value for desflurane, electricity, and beef.

2.6. Model interpretation

Interpreting GBM model results can be challenging because the model consists of a large number of decision trees with high interaction depths in each ensemble. To enhance our understanding of the relationship between predictor and outcome variables, we employed two widely used tools: relative influence and partial dependence of predictor variables (Friedman, 2001).

First, we estimated the relative influence of predictor variables. The influence of a predictor variable is calculated by the overall improvement in squared error when that variable is selected for splitting in the GBM model (Friedman, 2001). This improvement is then averaged across an ensemble of decision trees within the model (Friedman, 2001; Natekin & Knoll, 2013). The relative influence is defined as the standardized importance of a predictor variable relative to all other predictor variables, ensuring that the sum of relative influence across all variables equals 100% (Natekin & Knoll, 2013). By exploring the relative influence of predictor variables, we can gain deeper insights into the correlation between outcomes and influencing variables (Yang et al., 2020). Specifically, a higher value represents higher importance of the variable in the GBM model’s predictions.

Partial dependence is another valuable tool for interpreting the effects of predictor variables on outcome variables. It predicts the outcome variable across all values of a selected predictor variable while holding all other predictor variables at their sample mean values (Natekin & Knoll, 2013). This approach enhances our understanding of how individual predictor variables influence the outcome variables.

2.7. Greenhouse gas emission

We estimated greenhouse gas (GHG) emissions for each outcome (desflurane, electricity and beef consumption) using both the observed and imputed consumption data according to established methods and emission factors outlined in the Greenhouse Gas Protocol, the US Environmental Protection Agency, and published literature (GHG Protocol, 2017; Liu et al., 2015; USEPA, 2023). Overall, the method we used to estimate GHG emissions can be summarized using the following equation ( 1 ):

$${Emission}_{i,j}={Consumption}_i\times {EF}_{i,j}$$ ( 1 )

where i represents outcome variable i, and j indicates the type of GHG, including carbon dioxide, methane, nitrous oxide and other greenhouse gases. The emissions from non-CO₂ greenhouse gases were converted to CO₂ equivalent (CO₂e) values using 100-year Global Warming Potential (GWP) published by IPCC fifth assessment report (IPCC, 2015).

2.8. Uncertainty analysis

To evaluate the uncertainty in the GBM models, we employed a bootstrap sampling approach. Specifically, we randomly selected 90% of the observed data 100 times for the three outcome variables. Applying the resampled datasets, we constructed 100 GBM models for every outcome variable and utilized these models to make predictions for consumption. Subsequently, we computed 95% confidence intervals (95% CI) for the estimated consumption values of desflurane, electricity, and beef. Furthermore, we estimated the total GHG emissions along with their corresponding confidence intervals based on the consumption results.

3. Results

The final GBM models used the hyperparameters summarized in Table 1 and between 4% to 21% of the originally considered sets of predictor variables, depending on the outcome. Final models had excellent predictive ability for electricity and beef, with test R² of 0.82 for each, using 178 and 44 predictor variables, respectively. However, the final model for desflurane had less predictive ability, with test R² of 0.51 using 36 predictor variables. This less optimal performance may be due to the considerable heterogeneity of desflurane consumption compared to the other two outcomes (e.g., coefficients of variation were 1.99 for desflurane vs. 1.27 for electricity and 0.96 for beef in observed data) or because desflurane had the smallest number of observations (N=125), hampering the development of an accurate GBM model to predict missing values in other hospitals.

Table 1.

Selected hyperparameters for different outcome variables

Hyper parameters	Desflurane	Electricity	Beef
Number of trees	2000	2000	2000
Interaction depth	3	3	8
Observation in terminal nodes	5	2	10
Shrinkage rate	0.01	0.1	0.1
Subsampling	0.8	0.8	0.8
CV folds	5	5	8
Best MSE	1.32	0.25	0.21
Model performance as assessed by R2 from
Training data	0.84	1.00	0.95
Cross-validation in training data	0.30	0.81	0.79
Test data**	0.51	0.82	0.82

Notes: *Analysis data was divided into 70% training and 30% test datasets.

^**Test data R2 is typically the most generalizable measure of performance.

Figure 1 shows the relative influence of the top 20 predictor variables for each outcome. For desflurane, dollars spent on desflurane was the most important predictor, contributing to 32% of the variation in consumption of desflurane, followed by the number of emergency department (ED) visits (10%) and whether the food service shut down during the COVID-19 pandemic (10%). For electricity, the most important predictor was building area (39%), followed by: annual water consumption (22%), and number of operating rooms (9%). For beef, variables with the largest influence were solid waste weight (30%), dollars spent on food and beverages (25%), and number of staffed beds (11%).

Figure 1.

Relative influence of predictor variables on the outcome variables

Notes: The abbreviation names of predictor variables are explained in the appendix Figure A4.

Figure 2 shows partial dependence plots of the modeled relations between each outcome and its top 5 predictor variables. Consumption of desflurane was positively associated with desflurane expenditure, number of ED visits, annual OR procedures, and number of customized OR procedure packs prepared for specific procedures and negatively associated with food service shutdown during the COVID-19 pandemic. Hospitals that shut down their food services may have been more likely to postpone non-emergent surgeries and thus reduced the amount of desflurane needed (ACS, ASA, AORN, AHA, 2020). It was also possible that hospitals were more attentive to their resource usage during the pandemic, and prioritized alternative ways to provide anesthesia that were more cost-effective (Dohlman et al., 2020). The relationship between electricity usage and natural gas usage followed a nonlinear pattern. Initially, as natural gas consumption increased, electricity consumption also increased, reaching a peak at 245,723 Metric Million British Thermal Unit (MMBtu). However, beyond this point, further increases in natural gas consumption were accompanied by a decrease in electricity consumption. This suggests that natural gas can substitute some of the electricity consumption but not all of it, given that hospitals rely on electricity-powered services and devices. The consumption of beef showed a positive association with the top five predictors: solid waste weight, spending on food and beverages, number of staffed beds, number of licensed beds, and patient days.

Figure 2.

Partial dependence plots for top 5 influencing variables in each outcome

Notes: Please note that the y-axis of each outcome variable is log-transformed. The relative influence (RI) of the corresponding predictor variable on the outcome variable is shown on the lower right.

Considering both observed and imputed data, electricity usage in hospitals accounted for the highest amount of greenhouse gas (GHG) emissions, totaling 2,391,000 (95% confidence interval: 2,390,000, 2,393,000) metric tons of CO₂ equivalent (MTCO₂e). Beef consumption contributed 577,000 MTCO₂e (570,000, 584,000), while desflurane usage contributed 29,000 MTCO₂e (28,600, 29,600). On average, each hospital’s annual electricity consumption contributed 8,467 MTCO₂e (7,095, 9,839), beef consumption contributed 2,058 MTCO₂e (1,831, 2,286), and desflurane consumption contributed 99 MTCO₂e (73, 126).

Figure 3A provides estimates of GHG emissions using both observed and imputed data for electricity, desflurane, and beef consumption. The GHG emissions per hospital resulting from observed consumption of desflurane, electricity, and beef was higher than imputed estimates. The observed consumption of desflurane per hospital contributed 165 MTCO₂e (107, 223), whereas the predicted desflurane consumption resulted in GHG emissions of 54 MTCO₂e (49, 58) per hospital. Similarly, hospitals with observed electricity consumption exhibited average GHG emissions of 8,644 MTCO₂e (7,198, 10,090), which were 73% higher than the GHG emissions per hospital predicted with imputed consumption. The observed beef consumption accounted for 2,305 MTCO₂e (1,954, 2,655), which was approximately 30% higher than the GHG emissions estimated from imputed beef consumption.

Figure 3.

Estimated GHG emission after imputation for each outcome variable.

Notes: The middle lines inside the boxplots represent the median values, while lower bound and the upper bounds of boxplots represent 25^th and 75^th percentiles. The whiskers illustrate the emission values outside the inter-quantile range.

The differences in estimates of GHG emissions between observed and predicted consumption were primarily attributed to a higher missing rate in consumption data for smaller hospitals. Specifically, for desflurane consumption, the missing rate in small hospitals with fewer than 100 licensed beds was 69%, while the missing rate for larger hospitals (licensed beds ≥ 500) was 43%. Similarly, the missing rate in electricity consumption among small hospitals was 6%, compared to a 1% missing rate in large hospitals. Beef consumption was missing for 51% in the small hospitals compared with 39% in large hospitals. As a result, the GHG emissions estimated from the predicted consumption data were lower than the emissions calculated from the observed data.

To compare GHG emissions across hospitals of different sizes, we normalized the emissions by the number of licensed beds in each hospital (Figure 3). The GHG emission per licensed bed was found to be 31.8 MTCO₂e (29.0, 34.7) for electricity, 7.3 MTCO₂e (6.8, 7.8) for beef, and 0.4 MTCO₂e (0.3, 0.5) for desflurane. Normalization of GHG emissions by licensed beds made the distribution of predicted and observed emissions more similar for desflurane and reduced the spread in the distribution of emissions for electricity and beef. These variations in GHG emissions per licensed bed could also be influenced by a higher proportion of data imputed for smaller hospitals than for larger hospitals.

4. Discussion

To address the issue of missing data on GHG emission sources in a consortium of medical facilities, we developed an approach using GBM to predict hospital-level usage of desflurane, electricity consumption, and purchased beef. We imputed missing values for these GHG sources, and used observed and imputed values to calculate carbon emissions. Our results indicate that the GBM machine learning technique can accurately impute missing data for electricity and beef consumption, whereas the imputation accuracy was relatively lower for desflurane. It is possible that hospitals reduced their use of desflurane in response to sustainability initiatives (Zoghbi et al., 2022). Therefore, including additional variables related to sustainability practices may improve the model’s performance.

The consumption of electricity, beef, and desflurane in the 283 hospitals studied resulted in over 3 million MTCO₂e GHG emissions in 2020. These emissions were significant, equivalent to almost 45% of the national GHG emissions from the healthcare sector in Austria in 2014 (Weisz et al., 2020) and over 12% of healthcare emissions in the UK in 2019 (Tennison et al., 2021). Of the three sources, electricity dominated the carbon footprint, contributing to nearly 80% of the total emissions from these sources. Further analysis across the 283 hospitals revealed substantial variation in GHG emissions per licensed bed, suggesting there are potential opportunities for targeted interventions to reduce the carbon footprint, particularly in hospitals with emissions per licensed bed above the average. GHG emissions from observed data were found to be higher than those estimated from predicted data. However, the total GHG emission per capita of the three outcome variables estimated from observed data tended to be lower than the predicted GHG emission per licensed bed. We speculate that smaller hospitals, which often had more missing data, had a smaller total carbon footprint but were less efficient compared to larger hospitals.

This study also provides insights into the importance of data collection strategy for future missing data imputation. We found that predictor variables that reflect the operational volume, such as the number of ED visits, annual OR procedures, building area, number of operating rooms and staffed beds, were often the most influential factors that positively influenced resource consumption. Not surprisingly, amount spent on desflurane and on food and beverages were top predictors in imputing missing values in desflurane and beef consumption. All these variables might be considered a priority for hospital data collection to be able to estimate emissions, at least for these outcomes.

The strength of the GBM approach lies in its capability to make predictions using readily available variables. The primary objective of this analysis was to predict outcomes of interest, rather than to identify potential interventions or assess the impact of interventions aimed at reducing GHG emissions, for example of educational campaigns to reduce use of meat. We do suggest some operational practices to help healthcare facilities reduce their carbon footprint. First, we find that there was a positive association between electricity usage in hospitals and factors such as the size of the building area, annual water consumption, and the number of operating rooms. This suggests that larger hospitals tend to have higher electricity consumption, which can be attributed to various factors including increased lighting needs, the power requirements of medical devices, and water pumping usage. Therefore, efforts aimed at improving energy efficiency and implementing water conservation practices have potential to reduce the carbon footprint associated with electricity consumption in hospitals (Borges de Oliveira et al., 2021). Additionally, implementing mitigation strategies that focus on reducing climate impacts of electricity consumption can be effective in decreasing the carbon footprint of healthcare facilities. This can be achieved by procuring a larger share of renewable energy (Lee et al., 2023) and transitioning from fossil fuels to electricity, leveraging ongoing grid decarbonization efforts (Bozoudis et al., 2022; CEQ, 2021; EIA, 2023; Tennison et al., 2021).

Second, this study revealed a positive relationship between the amount spent on desflurane and its consumption in hospitals. Implementing sustainable practices such as reducing the use of desflurane and promoting alternative anesthesia methods like regional and total intravenous anesthesia can significantly mitigate the climate impacts associated with desflurane (McGain et al., 2020). These practices can also generate cost savings due to the lower expenses associated with alternative anesthetics and anesthesia techniques (Kampmeier et al., 2021; Moody et al., 2020). Lastly, the positive correlation between beef consumption in hospitals and the amount of solid waste generated, as well as the amount spent on food also has implications for sustainable food services in hospitals. Hospitals can implement food catering practices that optimize menu planning, portion control and demand forecasting (Dhir et al., 2020; Principato et al., 2018). By adopting such practices, hospitals can reduce beef consumption and effectively decrease the environmental footprint associated with beef consumption and food waste (Hennchen, 2019; Tran et al., 2021). Other practices, such as introducing more plant-based meal options can also play a significant role in reducing beef consumption and its carbon footprint (Bassi et al., 2022). These sustainable practices can also yield cost savings in food services in hospitals (US BLS, 2023). By adopting these strategies, healthcare facilities can enhance their sustainability efforts and reduce their climate impacts.

The primary objective of this study was to develop a GBM framework for imputing missing data related to resource consumption at the facility level. GBM models were chosen for their exceptional performance in terms of accuracy, efficiency, flexibility, and ability to handle missing data (Antipov & Pokryshevskaya, 2020; Coffman et al., 2020; Friedman, 2001; Hastie et al., 2009; Natekin & Knoll, 2013; Touzani et al., 2018). Future research endeavors may provide useful information comparing and exploring other machine learning methods, such as LightGBM and XGBoost models (C. Chen et al., 2020; Ke et al., 2017), to further improve the performance of the imputation model.

This study has some limitations that need to be considered when interpreting the results. First, the dataset was based on self-reported survey data, which may be subject to reporting bias and inaccuracies. This could have influenced the quality of the data used for imputation, which in turn may have affected the accuracy of the imputed values. Second, the sample of hospitals used in this study may not be representative of all hospitals in the US. Specifically, in the PGH survey, 72% of the hospitals were from large hospitals with more than 100 licensed beds. This percentage was found to be 70% higher than the corresponding national statistics (42% large hospitals) reported by Homeland Infrastructure Foundation-Level Data in 2020 (HIFLD, 2020). As a result, the imputation results obtained in this study may not be generalizable to smaller hospitals in the US. This study focused on estimating GHG emissions in 2020, coinciding with the onset of the COVID-19 pandemic. Given the unprecedented shocks and disruptions in healthcare systems caused by the pandemic (Lal et al., 2021), the estimated GHG emissions may not be representative of years without a public health emergency. Furthermore, the discrepancy in data collection periods poses a challenge when comparing data across hospitals, especially during the pandemic. Standardizing the data collection period would ensure greater consistency and accuracy in assessing and benchmarking sustainability efforts among healthcare facilities.

Future plans are to utilize the imputation approach developed in this study to estimate other sources of GHG emissions in healthcare facilities and extend the analyses to non-pandemic years. While PGH provides a valuable platform for healthcare facilities in the US to estimate their carbon footprint, the significant amount of missing data limits its effectiveness in providing guidance for decarbonization strategies, in part also due to the voluntary nature of GHG emission reporting for healthcare facilities in the US (Eckelman et al., 2020; HHS, 2022). In contrast, the UK National Health Service (NHS) mandates the reporting of GHG emissions from hospitals, providing a more comprehensive and reliable dataset for analysis (NHS, 2022; Singh et al., 2022). Such efforts have provided rich data and evidence-based support for the UK to become the first country aiming to achieve net-zero healthcare emissions by 2045 (NHS, 2022). The US and countries worldwide can learn from the NHS by developing a mandatory data platform that facilitates hospitals collecting, reporting (NHS, 2022), analyzing healthcare climate performance, and ultimately contributing to the global effort to mitigate climate change.

5. Conclusion

In conclusion, the GBM method shows promise in imputing missing values for monitoring and reporting of GHG emissions from healthcare facilities. By utilizing larger and more comprehensive datasets, the GBM imputation method could improve the accuracy of imputing missing values determining resource consumption, thus providing a more comprehensive and precise representation of GHG emissions from healthcare facilities. Further research could explore the potential of this approach in other sources of GHG emissions within hospitals, informing policy and regulatory frameworks for monitoring and reporting GHG emissions.