1. Collect data where it is most useful

Data collection for health system decarbonization analysis should minimize unnecessary burden. This means using, or building on, existing data collections or collection infrastructure where possible and avoiding duplicate requests for the same data. Manual data collection from a large number of organizations should be a last resort, and only where there is substantial benefit in terms of coverage or representativeness of resulting estimates. Analysts should prioritize additional data collection where it addresses gaps in emissions coverage, improves estimates of major sources, or supports improved analysis and monitoring of decarbonization interventions.

Across existing systems, automated reporting and manual collection, use consistent definitions, units, reporting periods and identifiers. Document differences between sources and any adjustments needed to make the data comparable.

Fig. 1 outlines three tiers of data collection methods in order of preference. Where existing data are sufficiently reliable and cover the required activities and reporting period, they should be used rather than re-collected. Use automated reporting systems where they already exist, or consider developing them if they will provide long-term benefits. For example, in a system where many hospitals already have sophisticated building management systems, consistent data could be extracted and collected with minimal human interaction, although establishing this data stream may require significant resources.

Where new data collections are required, they should use existing reporting infrastructure where possible, integrating with existing processes within health systems. The addition or refinement of a few key metrics within an existing collection could add significant value with limited additional reporting burden. For some important emissions sources, supply to a health system may be dominated by a small number of suppliers. Collecting data from a handful of suppliers may be much more efficient than collecting from hundreds of separate organizations.

Existing financial and procurement records held by ministries of health and provider organizations can provide a starting point for spend-based analysis. Work with finance, procurement and estates teams and those responsible for national health information systems, including routine health information systems. National statistics offices and other government departments, such as finance, energy and transport ministries, may also hold data relevant to the analytical task. Shared facility identifiers, definitions and reporting periods can help link service activity, resource use and emissions. The WHO strategy for optimizing national routine health information systems (2) and SCORE technical package (3) provide resources for strengthening data systems and their use in planning and decision-making.

Before establishing a new manual collection, weigh the collection and reporting effort against its expected value, including uses beyond emissions monitoring. Focus on priority data needs and minimize reporting burden to support reliable, complete responses. Minimizing reporting burden not only saves resources within health organizations, but maximizes data quality by focusing efforts on high-quality and high-coverage data in the most important areas.

Different datasets may follow different routes. Use automated access or collection where it is feasible and proportionate, and check the suitability and quality of data obtained through each route.

2. Use appropriate platforms and technology

Technology requirements are likely to change over time as data granularity and coverage increase. Initially, the focus is likely to be on estimating emissions at national level, ensuring good coverage of relevant activities and identifying hotspots in a given system. Emissions from many sources can be estimated based on high-level assumptions or national-level input data. At this stage, the volume of data required may be relatively small and a well-constructed spreadsheet may be sufficient.

As data become more detailed and modelling expands to improve estimates in priority areas or at subnational level, data volumes can increase substantially and become difficult to manage reliably within a single spreadsheet. Where data coverage is incomplete, all assumptions should be agreed through appropriate governance processes and recorded for use in calculations. For example, a health system with 200 hospitals and 2000 primary care sites, assessed across 50 emission categories and 10 reporting years, might require 1.1 million separate emissions estimates, before accounting for the multiple inputs and calculations underlying each estimate. Matrix manipulation of multi-regional input-output tables containing more than ~50 country-sector pairs cannot be performed in spreadsheets due to calculation limitations.

As calculation requirements grow, structured databases and scripted workflows may be required, with clear approaches to linking activity and spend data with emission factors and combining figures from different organizations and emission sources. Data can be processed within databases using SQL scripts or stored procedures, or through scripts written in R or Python. These approaches can also support modelling to fill data gaps. Analysts should use technologies that are already commonly used within their health system where possible and avoid proprietary solutions, to improve resilience and portability.

Unique identification codes should be used to allow easy linking between datasets. These should align with codes used elsewhere in the health system, such as existing organization codes, but may need to be developed in other areas. Linking datasets on names alone should be avoided where possible, as there is a high chance of incomplete matching which could lead to incomplete coverage of resulting estimates.

Expenditure-based methods are likely to remain important for supply chain emissions where product-specific emission factors are unavailable. However, as the availability of these factors increases, it may lead to further order-of-magnitude increases in the volume and diversity of data used for emission estimates, given the complexity and diversity of health system supply chains. At this point, automated data ingestion and manipulation techniques, including use of application programming interfaces where available, will become increasingly necessary. Developing and maintaining these processes requires expertise that may already exist within health system data teams.

Health systems must retain ownership of the analysis and the capability to scrutinize, update and use the underlying data and models. Where possible, common calculations should be undertaken once at national level to reduce duplication, improve consistency and make efficient use of available expertise. Teams can also draw on shared international resources to reduce duplication of analytical effort. These include the Alliance for Transformative Action on Climate and Health (ATACH) Country Documents Repository for published country assessments and plans, the ATACH Resource Repository for guidance and tools, and the Lancet MedZero for emissions estimates specific to health care (4–7).

Short-term external support may be needed where worthwhile model improvements exceed the organization’s available skills or resources. If this route is taken, national-level coordination is likely to provide significantly better value than individual organizations separately procuring solutions. Such support must build rather than replace health system capability, with knowledge transfer and documentation integral to the work. Health system teams must retain access to the data, code and methods needed to maintain and develop the analysis over time. Without knowledge transfer, health systems risk dependence on one-off, for-profit assessments that cannot be scrutinized, updated or applied.

3. Move towards use of automated pipelines and version control

Version control for code and data supports reproducibility as models develop, particularly when estimates inform publications or national reports. Where possible, use reproducible analytical pipelines (8) alongside code version control, for example using Git.

Extensive quality assurance checks should be implemented (Fig. 2). When developing a hybrid model, it is particularly important to keep track of which parts of the scope are covered by which estimates, to avoid gaps or double counting. It is very easy for certain categories of activity or organizations to be accidentally omitted from final estimates, especially as hybrid models are likely to require different levels of granularity for different emissions areas, and because input data may change in form or coverage over time. Both automated checking and manual inspection of results are required to minimize the chance of this kind of error.

Where both data and methodologies are improving, reconciling estimates made at different time points can be challenging. However, being able to explain changes to estimates is vital to retain confidence in the modelling. Ideally, this is achieved through full reproducibility: the ability to rerun the model at a later date and produce the same outputs as previously. Reproducibility requires retaining the code, input data, assumptions and software environment used for each run.

This could be achieved through full code and data versioning, for example through containerization which provides a snapshot of the codebase at the time of a model run. The model configuration, modelled estimates, assumptions and complete outputs must also be retained. Where this is not possible, extensive documentation of the configuration of the model at the time of running, including code version, data inputs, any modelled estimates or assumptions, and complete outputs, should be clearly recorded and saved.

Models should support consistent estimates across reporting years so that progress can be assessed fairly. Use appropriate data and assumptions for each year, and retain dated versions of input data, mappings, emission factors and model configurations. When methodological improvements affect comparability, recalculate earlier estimates where feasible or explain breaks in the time series. Preserve previously published results and document revisions, distinguishing changes in methods or coverage from changes in activity or emissions intensity.

Where data support both emissions estimates and other indicators, use pipelines drawing on the same controlled source data to keep reporting consistent across all products.

4. Present data to support decision-making

Analysis should support specific decisions, such as identifying priorities, comparing interventions, planning investment and assessing implementation and progress. The data underpinning emissions models can also inform national and subnational decisions outside annual emissions reporting cycles. For example, regular monitoring of hospital energy use can identify waste, inefficiencies and opportunities to save energy. The WHO digital implementation investment guide (Fig. 3) illustrates how quality data can inform decisions and action, with their use generating further information needs (10).

Develop carbon and non-carbon indicators to monitor progress, such as energy use or the proportion of the fleet that is electric. Annex A of the NHS England Green Plan guidance (11) provides specific examples which might apply to other health systems.

Metrics and data should be incorporated into existing tools where possible, to ensure they are easily available to decision-makers. Ideally, key metrics should be integrated into existing performance frameworks for organizations, with more detailed data available to those in specific roles such as estates, energy or fleet managers.

Interpret decarbonization results alongside access, quality, equity and other relevant health system outcomes. WHO’s health system performance assessment framework (12) can help relate indicators to wider policy goals. For example, lower hospital activity may reflect successful prevention or reduced access to care; its implications require additional evidence.