MCF classifier: Estimating, standardizing, and stratifying medicine carbon footprints, at scale
Funding information: SBRI Healthcare Programme (SBRIH18P2007), an Accelerated Access Collaborative initiative, championed by the Academic Health Science Networks (now known as the Health Innovation Networks).
Abstract
Aims
Healthcare accounts for 5% of global greenhouse gas emissions, with medicines making a sizeable contribution. Product-level medicine emission data is limited, hindering mitigation efforts. To address this, we created Medicine Carbon Footprint (MCF) Classifier, to estimate, standardize, stratify and visualize medicine carbon footprints.
Methods
We used molecular weight and chemical structure to estimate the process mass intensity and global warming potential of the active pharmaceutical ingredient in small molecule medicines. This allowed us to estimate medicine carbon footprints per dose, which we categorized into MCF Ratings, accessible via a searchable web application, MCF Formulary. We performed comparison and sensitivity analyses to validate the ratings, and stratification analyses by therapeutic indication to identify priority areas for emission reduction interventions.
Results
We generated standardized medicine carbon footprints for 2214 products, with 38% rated LOW, 35% MEDIUM, 25% HIGH and 2% VERY HIGH. These products represented 2.2 billion NHS England prescribed doses in January 2023, with a total footprint of 140 000 tonnes CO2e, equivalent to the monthly emissions of 940 000 cars. Notably, three antibiotics—amoxicillin, flucloxacillin and penicillin V—contributed 15% of emissions. We estimate that implementing the recommended 20% antibiotic prescription reduction could save 4200 tonnes CO2e per month, equivalent to removing 29 000 cars.
Conclusions
Standardized medicine carbon footprints have utility in assessing and addressing the carbon emissions of medicines, and the potential to inform and catalyse changes needed to align better healthcare and net zero commitments.
What is already known about this subject
- Medicines contribute significantly to health system carbon emissions.
- Emissions related to medicines are often estimated using economic modelling as product-level carbon footprint information is limited.
- Literature searches found product-level carbon footprinting studies for specific medications but no standardized estimates or comparisons across small molecule medicines.
What this study adds
- This study introduces a standardized method to estimate the carbon footprint per dose of small molecule medicines.
- This study presents a landscape analysis of medicine carbon footprints revealing potential widespread opportunities for carbon-informed prescribing.
- This study identifies antibiotics as a major contributor to medicine carbon emissions, and advocates for antimicrobial stewardship as a strategy to reduce antimicrobial resistance and greenhouse emissions.
1 INTRODUCTION
The global climate crisis is one of the defining challenges of our time, with consequences for ecosystems, economies and human well-being. The urgency of addressing climate change is increasingly evident, as extreme weather events and disruptions to natural systems threaten the stability of our planet and our health. Healthcare has a unique responsibility as it must simultaneously manage the acute and chronic health challenges caused by climate change and reduce its contribution to the climate crisis.1, 2
The primary driver of climate change is the increase in greenhouse gas (GHG) emissions, which trap heat in the atmosphere, leading to a warming effect known as the greenhouse effect.3 Major GHGs include carbon dioxide (CO2), methane (CH4) and nitrous oxide (N2O).3 Human activities, particularly burning fossil fuels for energy, are the largest sources of these emissions, making them anthropogenic GHG emissions. A carbon footprint quantifies the total GHG emissions caused by an individual, organization, event or product, typically expressed in carbon dioxide equivalents (CO2e) to account for the different global warming potentials (GWP) of various gases.4 To measure these emissions, economic input–output models, which estimate emissions based on economic activity and expenditure data, are often used.5 This approach provides an accessible method for calculating emissions when direct measurement data is unavailable, serving as a general method to approximate the carbon footprint of various sectors and activities.6 However, it is not suitable for more granular comparisons, for example to compare products. To comprehensively assess the environmental impact of a product, process or service, life-cycle assessment (LCA) is considered the gold-standard. LCA is a systematic method that evaluates the environmental aspects and potential impacts from raw material extraction through production, use and disposal.7 This cradle-to-grave approach provides a holistic view of the product's environmental footprint, but it is currently time-consuming, costly and not standardized.
Healthcare contributes approximately 5% of global emissions, with medicines being a significant contributor, accounting for 25% of the NHS carbon footprint.8 Calculating pharmaceutical emissions typically relies on input–output economic modelling, estimating carbon emissions based on expenditure, because product-level carbon emission data for medicines is rarely accessible.8 Two exceptions are anaesthetics and respiratory inhalers.8-10 Anaesthetics emissions are quantified through direct measurement of the gas volume and the specific gas's GWP.9 Several anaesthetics have GWPs much greater than that of carbon dioxide (CO2e >> 1).11 For inhalers, attention has focused on quantifying and reducing emissions of propellants used to deliver medication into the lungs, some of which have high carbon footprints.10 These examples demonstrate the value of product-level medicine carbon footprints for identifying, prioritizing and reducing pharmaceutical emissions. The approach will be essential for the many health systems committed to achieving net zero emissions.12 However, the methods employed to date are not widely applicable because gases and propellants are rarely used in medicines.
About 90% of medicines are small molecules characterized by low molecular weight (typically <1000 Da).13 These medicines are important in most therapeutic areas, and include common drugs such as analgesics, antihypertensives and antibiotics. Due to their high volumes and clinical significance, the pharmaceutical industry has dedicated substantial effort to enhance the efficiency and sustainability of the development and production of small molecule medicines.14
Central to this has been the development of metrics to measure the efficiency and environmental impact of chemical reactions used in the manufacture of the active pharmaceutical ingredient (API). Among these metrics, process mass intensity (PMI) has become an industry standard.15, 16 PMI is derived by measuring the total mass of materials required to make 1 kg of API, including reagents, reactants, solvents and catalysts, and is dependent on the molecular weight, molecular structure and complexity of the synthetic process.
Initially, PMI served primarily as a tool for post-synthesis assessment, enabling the measurement of the environmental impact of drug manufacturing. However, its role has evolved, with PMI now playing a pivotal role in greener drug design. PMI prediction models have become foundational to optimizing the efficiency and resource-effectiveness of chemical synthesis processes in drug development, providing guidance for selecting the most eco-efficient synthetic pathways.15-18
PMI is strongly correlated with global warming potential, and API production accounts for the majority of carbon emissions associated with small molecule medicines (typically around 75–90%).15, 18 By comparison, formulation contributes ~10–15% while packaging and transport make small contributions.15, 18
Thus the carbon footprint of API production can be considered a practical and informative surrogate for the overall footprint of a medicine. This parallels inhalers, where propellants are used as surrogates to compare the relative environmental impacts of the entire inhaler.
Here, we have harnessed the models and metrics developed to measure the environmental impact of new small molecules to build Medicine Carbon Footprint (MCF) Classifier, a suite of applications to estimate and classify the carbon footprint per dose of small molecule medicines, at scale. We have used the generated data to analyse carbon emissions across medicines and therapeutic indications, and to show how such data can be used to identify priority areas for carbon-reduction interventions.
2 METHODS
2.1 Overview
MCF Classifier is a suite of data and applications used to generate, classify, integrate and visualize the carbon footprints of medicines. It includes five components as shown in Table 1.
| MCF Classifier | |
|---|---|
| MCF Method | An AI-driven, semi-automated pipeline to calculate the carbon emissions of medicines in gCO2e. |
| MCF Warehouse | A cloud store of all input and calculated data, optimized for rapid analytics across all medicines and ratings. |
| MCF Ratings | A colour-coded, log-based system that classifies carbon emissions into one of four emission categories: LOW, MEDIUM, HIGH, VERY HIGH. |
| MCF Products | Medicine products for which MCF Method is used to calculate the gCO2e per dose and MCF Ratings have been applied. |
| MCF Formulary | A searchable web-application to visualize MCF Ratings by therapeutic indication. |
2.2 MCF Method
In small molecule drug development, PMI prediction is used to evaluate the efficiency of different synthetic pathways and is strongly correlated with GWP.15, 18 Here we have used these principles to build an automated nine-step pipeline to estimate the PMI of small molecule medicines (Steps 1–6),16, 19, 20 which we input into a linear regression model to estimate the GWP of API production (Step 7)15 and the carbon emissions per dose (Steps 8–9). We call this pipeline MCF Method (Table 2).
| Step | Description | Method |
|---|---|---|
| 1 | Identify API (active pharmaceutical ingredient) | Data extraction21 |
| 2 | Identify SMILES (Simplified Molecular-Input Line-Entry System) | Data extraction22 |
| 3 | Calculate SAscore (synthetic accessibility score) | QSAR (quantitative structure–activity relationship) model19 |
| 4 | Calculate molecular complexity | Linear regression model20 |
| 5 | Identify molecular weight | Data extraction22 |
| 6 | Calculate PMI (process mass intensity) | Linear regression model16 |
| 7 | Calculate GWP (global warming potential) | Linear regression model15 |
| 8 | Identify dose | Data extraction21 |
| 9 | Calculate gCO2e per dose | Output Step 7 × output Step 8 |
The inputs for Steps 1, 2, 5 and 8 are fixed or empirical values. We used ETL (Extract, Transform, Load) pipelines to programmatically upload these data to the MCF Warehouse, a cloud-based architecture23 built using Python, SQL and Terraform.21, 22, 24-27 For Step 3 we evaluated several approaches and selected the model developed by Ertl et al.19 to calculate the synthetic accessibility score (SAscore), as it was automatable, scalable and shown to be well correlated with other models.19, 28 The SAscore is an estimate of the ease of synthesis of a molecule and is generated from the API represented as a SMILES (Simplified Molecular-Input Line-Entry System) string using a quantitative structure–activity relationship (QSAR) model.19 For Step 4, the SAscore is converted to a molecular complexity score using a linear equation from Sheridan et al.20 For Step 6, the molecular complexity and molecular weight are used to calculate PMI using the SMART-PMI model of Sherer et al.16 For Step 7, the linear regression model of Jimenez-Gonzalez et al.15 is used to estimate the GWP from the PMI. The output of Step 7 is multiplied by the dose (Step 8) to give the carbon footprint in gCO2e per dose (Step 9). Further details are provided in the Supporting Information.
2.3 MCF Method validation
There is very limited accessible GWP or gCO2e data for medicines, as such information is considered proprietary by medicine manufacturers. Building MCF Method using the models developed and validated by medicine manufacturers compensates for this, to some extent, but we still sought to compare our outputs with orthogonal data.
To validate our GWP calculations we used data from the Association of the British Pharmaceutical Industry (ABPI) Blister Pack Carbon Evaluation Tool.29 The ABPI tool developed a statistical model to estimate the GWP of APIs, using data shared confidentially by manufacturers. The individual data points were not made available; therefore, we compared the average GWP and distribution of GWPs between MCF Method and the ABPI tool.
To compare the final gCO2e per dose we used data provided to one of the authors (M.S.) from the manufacturer. M.S. contacted the manufacturers of 20 commonly prescribed products. Only one manufacturer supplied gCO2e per dose data from internal LCA performed by a third party for two of the products, omeprazole and rosuvastatin. No information about methods or LCA inclusion criteria were provided.
2.4 MCF Ratings and sensitivity analysis
The gCO2e per dose were stratified into four categories, LOW, MEDIUM, HIGH, VERY HIGH, following a log10 scale. The logarithmic bounds were chosen to give an even distribution across products (Table 3).
| MCF rating | gCO2 per dose |
|---|---|
| LOW | <10 |
| MEDIUM | ≥10–100 |
| HIGH | ≥100–1000 |
| VERY HIGH | ≥1000 |
We performed sensitivity analyses to investigate the impact of model uncertainty on ratings, by adjusting the carbon footprint per dose values up and down by 5%, 10% and 15% and calculating the number of products for which the MCF Rating category changed.
2.5 MCF Formulary
To provide a user-friendly and practical interface to explore MCF Ratings, we built the MCF Formulary, informed by user research and tested at concept, design, prototype and beta stages. MCF Formulary is an interactive web application with a Remix web framework frontend to ensure accessibility for users with older browsers.30 The data elements are housed in the MCF Warehouse and accessed via an API layer built in Python, which allows for fast database retrieval, scalable to many concurrent users.
MCF Formulary is free to access, for non-commercial uses, via a simple sign in, sign out functionality.31 The login and account features follow the recognized authentication and authorization standards.31, 32
When a user searches for a medicine in MCF Formulary, the encoded entry is retrieved from MCF Warehouse and is displayed alongside entries for all similar medicines (as defined in the British National Formulary). The most prescribed MCF Product for each medicine and its corresponding MCF Rating is presented with an expandable dropdown to view all other products for that medicine and their MCF Ratings ordered from most to least prescribed. The streamlined presentation of the most prescribed MCF Product for similar medicines and their MCF Rating allows the user to quickly compare the carbon footprint of medicines which may be prescribed in a similar context. The user can then drill down into each medicine to find the specific product they are interested in and see its MCF Rating.
The MCF Formulary was made live at https://formulary.yewmaker.com/ on 4 March 2024.33
2.6 MCF Products
MCF Products are the medicines for which we have calculated per dose emissions. For this study, we applied the following product selection filters.
2.6.1 Automation filters
Only products with a single API were included to facilitate automation of MCF Method Steps 1, 2 and 5. Only tablet or capsule products for which the API quantity was given in units convertible to grams were included to allow automation of Step 8.
2.6.2 Outlier filter
SAscore and molecular weight have a known, expected correlation such that larger molecules are typically harder to synthesize than smaller molecules.19 This correlation was validated in our data. We also observed a bias in the SAscore model whereby very small molecules were given a very high SAscore, which is most likely caused because such molecules rarely appeared in the SAscore training data. We removed these outliers to minimize bias by implementing threshold values. Specifically, any active pharmaceutical ingredient with an SAscore above 5 and a molecular weight below 150 was removed.
2.6.3 Product filters
Included products were required to have Dictionary of medicine and devices (dm + d) Virtual Medicinal Product (VMP) Systematized Nomenclature of Medicine Clinical Terms (SNOMED) code,21 a British National Formulary (BNF) code,24 and a BNF to SNOMED mapping entry.25 We also used filters to ensure each unique product (API + dose) was only represented once in the dataset irrespective of the number of different product manufacturers.
The dm + d VMP code was used as an entry point into the dm + d system providing encoded information on the API, drug form and dosage. The BNF to SNOMED mapping was used to link the BNF and dm + d systems for each product.
2.6.4 Therapeutic classification filters
The BNF code was used in the stratification analyses by therapeutic indication. The BNF uses a hierarchical classification system to group medicines based on their therapeutic use(s). At the highest level, medicines are divided into chapters, with each chapter corresponding to a broad therapeutic area or medical specialty. Each chapter is further divided into sections which are further divided into paragraphs. We simplified and systematized the chapter names for consistency and coherence. The therapeutic categories we used and their corresponding BNF chapter names are given in Table 4.
| BNF code | BNF chapter name | MCF category name |
|---|---|---|
| BNF 1 | GI System | Gastro-intestinal |
| BNF 2 | Cardiovascular System | Cardiovascular |
| BNF 3 | Respiratory | Respiratory |
| BNF 4 | CNS | Neurology |
| BNF 5 | Infections | Infections |
| BNF 6 | Endocrine | Endocrine |
| BNF 7 | Obs, Gynae & Urinary-tract | Gynae & urology |
| BNF 8 | Malignant disease and Immunosuppression | Oncology |
| BNF 9 | Nutrition and Blood | Haematology |
| BNF 10 | MSK & Joint diseases | Rheumatology |
2.7 Stratification analyses
To investigate how medicine carbon footprints are distributed by therapeutic indication we generated rug plots of MCF Ratings for the first 10 BNF therapeutic categories. Each tick along the x-axis in the rug plot represents a single product and is coloured by its MCF Rating. Within each therapeutic category, products are ordered by their MCF Rating.
To estimate the total monthly carbon emissions of prescribed doses in NHS England of MCF Products, we accessed the total quantity of doses for primary26 and secondary27 care for 1 month, January 2023, from the NHS Business Services Authority, and multiplied the quantities of doses by the gCO2e per dose. To make the data more accessible, we estimated the equivalent monthly car emissions of the average car (unknown fuel type) using the emissions data from the Department for Energy Security and Net Zero (DESNZ) and England mileage data from the Department for Transport (DfT).34, 35
2.8 Role of the funding source
The funders of the study had no role in study design, data collection, data analysis, data interpretation or writing of the paper.
3 RESULTS
Using MCF Method and the selection filters, we generated medicine carbon footprints for 2214 products (referred to as MCF Products).
We compared GWPs generated with MCF Method and the ABPI tool.29 The distribution implied by the boundaries used in the ABPI tool is similar to the MCF GWP distribution (Figure 1A), and their average GWP of 1500 kgCO2e/kg is similar to the average MCF GWP of 1300 kgCO2e/kg. We also compared the gCO2e per dose calculated with MCF Method for omeprazole and rosuvastatin, with data supplied by the manufacturer from LCA. The results were comparable. As anticipated, the emissions per dose calculated by MCF Method, which are based on the API production emissions, were slightly lower (~10%) than from LCA, as the latter also includes emissions from other components such as formulation (Figure 1B).
We categorized the 2214 product per dose emissions into LOW, MEDIUM, HIGH and VERY HIGH MCF Ratings. A total of 845 (38%) medicines were rated LOW (<10 gCO2e per dose), 770 (35%) were rated MEDIUM (≥10–100 gCO2e per dose), 557 (25%) were rated HIGH (≥100–1000 gCO2e per dose) and 42 (2%) products were rated VERY HIGH (≥1000 gCO2e per dose) (Table 5, Figure 2A).
| MCF Rating | gCO2e per dose | Number of products | Percentage of products |
|---|---|---|---|
| LOW | <10 | 845 | 38% |
| MEDIUM | ≥10–100 | 770 | 35% |
| HIGH | ≥100–1000 | 557 | 25% |
| VERY HIGH | ≥1000 | 42 | 2% |
We performed sensitivity analyses to explore how changes in the per dose calculations impact MCF Ratings. Overall, impacts were modest. For example, a 10% increase in the gCO2e per dose impacted less than 4% of products, with 39 changing from LOW to MEDIUM, 33 from MEDIUM to HIGH, and 10 from HIGH to VERY HIGH (Table 6).
| Change in gCO2e per dose | Number of products that change category | ||
|---|---|---|---|
| Low → Medium | Medium → High | High → Very high | |
| 5% increase | 23 (2%) | 22 (2%) | 5 (0%) |
| 10% increase | 39 (4%) | 33 (4%) | 10 (1%) |
| 15% increase | 54 (6%) | 46 (5%) | 13 (2%) |
| Very High → High | High → Medium | Medium → Low | |
|---|---|---|---|
| 5% decrease | 5 (12%) | 12 (2%) | 15 (2%) |
| 10% decrease | 6 (14%) | 30 (5%) | 32 (4%) |
| 15% decrease | 11 (26%) | 55 (10%) | 45 (6%) |
To provide a user-friendly interface to explore MCF Ratings, we built the MCF Formulary, an interactive web application (Figure 3). The MCF Formulary shows the MCF Rating of the searched medicine and the ratings of other medicines used in similar therapeutic indications as defined by the BNF. For each medicine the most prescribed product based on NHS England prescriptions in 2023 is shown by default (e.g., valsartan 160 mg in the example in Figure 3). Ratings of other available doses of the medicine, ordered by number of NHS England prescriptions in 2023, are accessible via a dropdown (e.g., valsartan 80 mg, valsartan 40 mg, valsartan 320 mg).
In the 4 months since its launch, MCF Formulary has had 100% uptime, 11 000 page views and 2700 active users. There have been no reported issues with accessing MCF Formulary or in understanding how to use it.
We next considered the attributable emissions of prescribed doses of the 2214 MCF Products. Some 2.2 billion doses were prescribed in England in January 2023 with a total carbon footprint of 140 000 tonnes CO2e which is equivalent to monthly emissions of 940 000 cars.
We stratified the 2.2 billion doses by MCF Ratings. The majority of emissions were attributable to products with a HIGH MCF Rating. HIGH-rated medicines accounted for 25% of products and 70% of emissions (Figure 2B).
We generated rug plots to explore how the medicine carbon footprints are distributed by therapeutic indication (Figure 4). All indications include medicines with LOW, MEDIUM and HIGH ratings, suggesting opportunities for carbon-informed prescribing to reduce medicine emissions may exist in many clinical scenarios.
The median per dose footprint by therapeutic category ranged from 8 gCO2e to 219 gCO2e. Infections had the highest median per dose footprint, primarily due to antibacterial drugs (hereafter called antibiotics). Of emissions from products used to treat infections in January 2023, 86% were attributable to antibiotics (Figure 5A), and 60% from penicillins (Figure 5B) of which almost all were due to amoxicillin (49%), flucloxacillin (31%) or penicillin V (19%) (Figure 5C). Indeed, 15% of the emissions (2% of doses) for all 2214 MCF Products in January 2023 were due to these three medicines, equating to 21 000 tonnes CO2e, which is equivalent to monthly emissions of 140 000 cars. If the recommended 20% reduction in prescription of these antibiotics36, 37 could be achieved, it would save 4200 tonnes CO2e per month, equivalent to taking 29 000 cars off the road.
4 DISCUSSION
To our knowledge, this study is the first to generate standardized carbon footprints for small molecule medicines and the first analysis of medicine carbon footprints across therapeutic areas. The overall analysis of 2214 products revealed a broad range of carbon footprints per dose, reflecting inherent diversity in molecular complexity and dosing regimens.
To categorize and compare the medicine carbon footprints, we implemented a logarithmic based MCF Ratings system, which has several advantages. First, as our sensitivity analysis shows, it is robust to uncertainty in MCF Method and its underlying models. Second, the logarithmic scale ensures ratings have substantial, systematic, differences between categories: a typical LOW-rated medicine has a 10-fold lower footprint than a typical MEDIUM-rated medicine which has a 10-fold lower footprint than a typical HIGH-rated medicine. Third, the scale is designed to provide a balanced distribution of ratings across products. This enables standardized, comprehensive evaluations across medicines, aiding healthcare professionals and policymakers in making carbon-informed decisions. Finally, the ratings are designed to have a familiar colour-coded traffic-light format which our user testing showed enhanced accessibility and usability.
We believe the MCF Ratings could have utility in many varied analyses and applications. In this study, we stratified and compared MCF Ratings by quantity of doses prescribed and therapeutic indication.
Carbon emissions associated with prescriptions are strongly correlated with MCF Ratings, with three-quarters of emissions attributable to the quarter of products with the highest ratings. This suggested there may be high-volume, HIGH-rated MCF Products that could be prioritized for carbon reduction strategies. Further analyses confirmed this hypothesis and we show three antibiotics—amoxicillin, flucloxacillin and penicillin V—contributed 15% of prescription emissions due to MCF Products, in England in January 2023. Amoxicillin 500 mg capsules had the highest monthly emissions among the 2214 evaluated products, despite being only the 38th most prescribed product, and was responsible for 10 000 tonnes CO2e, equivalent to the monthly emissions of 69 000 cars.
The responsible use of antibiotics is an urgent global priority, with antibiotic resistance predicted to lead to 10 million annual deaths by 2050 if no action is taken.38 Current estimates suggest 20–40% of NHS antibiotic prescriptions are inappropriate, unnecessary or exceed recommended durations.39-41 In this study, we have estimated the carbon emissions associated with prevailing antibiotic prescribing practices. The potential convergence of net zero and antimicrobial stewardship goals offers an exceptional opportunity to simultaneously accelerate progress towards several critically important global objectives: combating antimicrobial resistance, reducing unnecessary medications, reducing healthcare costs and reducing healthcare emissions.
Our analyses by therapeutic indication revealed a consistent pattern, with medicines spanning LOW, MEDIUM and HIGH ratings in all 10 therapy areas evaluated. This suggests opportunities for carbon-informed prescribing that maintain or enhance clinical care and reduce emissions may be more widespread than previously recognized. This merits further exploration, with a nuanced approach to ensure different therapeutic contexts and uses of medicines are thoroughly and appropriately considered. Future research should also investigate how prescribers engage with MCF Ratings, interpret the implications, and how these insights lead to changes in prescribing behaviour and emission reductions.
While our study provides valuable insights into the carbon footprint of medicines and their potential implications for healthcare emissions, it has limitations. Most importantly, the per dose carbon emissions we have generated are predicted using chemical and pharmaceutical manufacturing models; they are not measured directly. Direct measurement, through LCA, is still uncommon for medicines, and the results are rarely disclosed. Furthermore, the absence of a standardized LCA for medicines complicates comparisons. LCAs are crucial for manufacturers to identify environmental hotspots and sustainable optimizations within their own production processes. However, they are currently not yet widespread or sufficiently standardized for benchmarking or carbon-informed medicine prioritizations.42 By comparison, the methodological consistency and scalability of MCF Classifier are distinctive strengths, enabling its application in benchmarking thousands of products.
Our comparative analyses were constrained by limited data availability from orthogonal models. Furthermore, all current models were trained on modest datasets. There is potential for improvement in these underlying models with inclusion of more extensive data, and greater transparency and data disclosure would be advantageous to the whole field. Despite these constraints, it was reassuring that MCF Method values were comparable to the orthogonally derived data, and the resulting MCF Ratings were the same. In our comparative analyses of PMI, GWP and gCO2e per dose, MCF Method outputs were slightly lower than the manufacturer data. This discrepancy is not unexpected, given the comparative GWP estimates were from drug development processes, which typically yield higher emissions than more optimized commercial manufacturing, and the manufacturers' gCO2e per dose were based on LCA data, whereas MCF Method is based on API production. Future versions of MCF Classifier will include data for other components, such as excipients and packaging, to provide comprehensive cradle-to-gate product carbon emissions data compliant with the GHG Product Protocol standard,43 PAS 205044 and ISO 14067 standard.45 These components typically make minor and similar contributions across different medicines, thus their addition will enhance our estimates while maintaining systematic comparability of the relative values, particularly the MCF Ratings, as supported by our sensitivity analyses.
In the current study we restricted our analyses to products containing a single API and to tablet and capsule formulations. This will have led to an underestimation of the total carbon emissions of therapeutic classes. For example, appreciable quantities of amoxicillin are given in liquid form by mouth, injection or infusion, or as tablets in combination with clavulanic acid. Combination medicines and other formulation types, such as liquids and injectables, are being integrated into the next version of MCF Classifier. We also plan to include other therapeutic modalities to which we can extrapolate our methods, such as monoclonal antibodies, proteins and oligonucleotides.
Another limitation of our current analyses is their focus on the NHS in England. While the results of the quantity analyses will vary across different healthcare systems, our analysis of pharmaceutical-related emissions has global applicability, as it encompasses many medicines commonly used worldwide. The wide distribution of carbon footprints across therapeutic areas is also likely to hold true in different health systems. Of note, we designed MCF Classifier as a modular software suite adaptable to any health system, and we are currently extending our datasets and analyses to other health systems.
In addition to our own work, we hope this study and the accessibility of MCF Ratings in the MCF Formulary will stimulate awareness, investigations, advancements and innovations to measure and mitigate the carbon emissions associated with medicines. This must become a priority objective if healthcare is to reach net zero, and to deliver on healthcare's foundational responsibility to improve health and do no harm.
AUTHOR CONTRIBUTIONS
Haroon Taylor and Nazneen Rahman undertook the study design, method development, data acquisition, data analysis and manuscript writing. Shazia Mahamdallie was involved in the study design, data review and manuscript review. Matthew Sawyer was involved in method development, data acquisition and manuscript review. All authors were involved in the revision of the manuscript and approved the final version.
ACKNOWLEDGEMENTS
This work was funded by the SBRI Healthcare programme, grant number SBRIH18P2007. SBRI Healthcare is an Accelerated Access Collaborative (AAC) initiative, championed by the Academic Health Science Networks (AHSNs), which are now known as Health Innovation Networks. The views expressed in the publication are those of the authors and not necessarily those of the SBRI Healthcare programme or its stakeholders. We thank Alaeddine Bouslama, Mohamed Dhia Wesleti, Federico Dionisi and Fatma Ben Yedder from Think-it (www.think-it.io) for data engineering assistance in building MCF Formulary. We thank Helena Traill and Charlie Port, from Nooh studio (https://noohstudio.com/), for assistance in MCF Formulary design. We are grateful to the members of the MCF Advisory group and the Sustainable Medicines Partnership for their thoughtful input and advice.
CONFLICT OF INTEREST STATEMENT
H.T. and N.R. are employees of and hold shares in YewMaker. S.M. is a former employee of and holds shares in YewMaker. N.R. is Non-Executive Director at AstraZeneca. M.S. provides paid educational consultancy on healthcare environmental sustainability.
Open Research
DATA AVAILABILITY STATEMENT
The MCF Ratings for all medicines included in this study are publicly available through the MCF Formulary at https://formulary.yewmaker.com/.33
