Health Data Research in Australia: Access and Five Safes
Health data research in Australia runs on a layered system: the Privacy Act 1988 and state health privacy laws set the legal floor, the Data Availability and Transparency Act 2022 (DATA Act) creates a national scheme for sharing Australian Government data, and agencies such as the Australian Bureau of Statistics (ABS) and the Australian Institute of Health and Welfare (AIHW) grant access through secure environments governed by the Five Safes framework. Researchers typically never download identifiable data. They work inside approved secure settings, and only checked outputs are released.
Why Australia matters now
Australia has some of the richest linked administrative health data in the world. Universal health insurance through Medicare, the Pharmaceutical Benefits Scheme (PBS), hospital morbidity collections in every state and territory, and long-running cancer and death registries give researchers population-wide coverage that few countries can match. Linking these datasets is how Australian researchers study drug safety, health service use and outcomes across entire populations.
The legal environment is also moving. Parliament passed the first tranche of Privacy Act reforms in late 2024, including a statutory tort for serious invasions of privacy and new transparency requirements for automated decision-making, with further reforms still under discussion. The DATA Act has been operating since 2022, and its statutory review is shaping how the scheme evolves. Genomics is scaling too: Australian Genomics has built a national network for clinical genomics research, and the federal government has announced plans for a national body, Genomics Australia.
For anyone building health AI in Australia, this combination creates a clear expectation. Sensitive data is to be analyzed in controlled environments, under documented approvals, with outputs checked before release.
How the Five Safes shape access in Australia
The Five Safes framework was developed by the UK Office for National Statistics and is now embedded in Australian practice. The ABS uses it to manage access to its DataLab environment, and the AIHW applies it to its own secure data access. The framework asks five questions about any research use of data.
- Safe projects: is the use appropriate and in the public interest?
- Safe people: are the researchers trained and authorized?
- Safe settings: does the environment prevent unauthorized use?
- Safe data: has the data been treated to reduce disclosure risk?
- Safe outputs: are the results non-disclosive before release?
The DATA Act builds the same thinking into law. Its data sharing principles cover the project, the people, the setting, the data and the outputs, closely mirroring the Five Safes. The scheme is overseen by the National Data Commissioner. Commonwealth bodies act as data custodians, researchers and organizations become accredited users, and accredited data service providers can perform complex services such as data integration and secure access on a custodian’s behalf.
Australia’s national research data infrastructure
Several public programs make up the backbone of Australian health data research. None of them is a single central database. Most are networks of custodians connected by linkage and secure access services.
| Component | Role | How researchers use it |
|---|---|---|
| Privacy Act 1988 and Australian Privacy Principles | National privacy law covering Commonwealth agencies and many private organizations | Sets the conditions for collection, use and disclosure, including research use of health information |
| State and territory health privacy laws | Govern state health departments and public hospitals | Apply to hospital and registry data held by the states |
| Data Availability and Transparency Act 2022 | National scheme for sharing Australian Government data | Accredited users request data from Commonwealth custodians under the data sharing principles |
| ABS DataLab | Secure analysis environment for integrated government data | Approved researchers analyze unit-record data remotely with outputs checked before release |
| AIHW | National health and welfare statistics agency and accredited integrating authority | Provides linked national datasets and secure access for approved projects |
| Population Health Research Network (PHRN) | National data linkage infrastructure with state linkage units | Links records across jurisdictions using privacy-preserving linkage methods |
| Sax Institute’s Secure Unified Research Environment (SURE) | Remote-access research environment for linked health data | Researchers work on linked data without downloading it |
| NHMRC National Statement on Ethical Conduct in Human Research | National ethics guidance | Sets the rules human research ethics committees apply, including for waivers of consent |
A pattern stands out. Linkage is separated from analysis. Linkage units see identifiers but not clinical content, and researchers see clinical content but not identifiers. This separation principle is one of the reasons Australian linkage has earned public trust, and it maps directly onto the architecture of a modern Trusted Research Environment (TRE).
The Sovereign AI angle
Australia’s infrastructure was designed for statistical analysis. The next wave of research demand is AI: training models on hospital records, imaging and genomic data, and validating them on Australian populations before they are used in care. That raises a sovereignty question. If a model is trained on Australian health data, where does training happen, who controls the resulting model, and can the data or model be moved offshore?
Sovereign AI answers that question by keeping both data and computation under national control. The strongest version of this is federation. Instead of copying hospital, registry and genomic data into one place for AI work, a federated TRE runs at each custodian. Approved code, including model training jobs, travels to the data, and only reviewed aggregate results or model parameters come back. Because data never leaves the source, the approach suits a country where data is held by Commonwealth agencies, eight state and territory health systems, and many research institutes.
Federation also respects Australia’s existing custodianship model. The custodian keeps legal responsibility and technical control, the Five Safes still apply to every project, and outputs pass through an automated airlock that enforces disclosure rules consistently. Residency is easier to demonstrate when data physically stays in Australian infrastructure. Our guide to health data residency requirements covers how residency rules differ across jurisdictions.
Indigenous data sovereignty
Sovereignty in Australia has a second meaning that any health data program must respect. Indigenous data sovereignty is the right of Aboriginal and Torres Strait Islander peoples to govern the collection, ownership and use of data about their communities. The Maiam nayri Wingara principles, developed by the Maiam nayri Wingara Indigenous Data Sovereignty Collective, and the international CARE Principles for Indigenous Data Governance (collective benefit, authority to control, responsibility and ethics) set out what that means in practice. Ethics guidance from the NHMRC also addresses research involving Aboriginal and Torres Strait Islander peoples.
Federated infrastructure helps here as well. When data stays with the organizations entrusted with it, including Aboriginal community-controlled health organizations where they hold data, governance decisions can be made by the right people and enforced technically, rather than relying on the policies of a distant central repository.
What good practice looks like
A national genomics or biobank program is the clearest example of where federation fits. Genomic data is large, uniquely identifying and costly to move, and genomic research needs clinical context held elsewhere. Programs overseas, including Genomics England and Canada’s CanPath, analyze sensitive data inside federated TREs so that approved researchers can combine genomic and clinical data without extracting either. The same pattern would let an Australian program connect genomic data with linked state hospital records while each dataset stays with its custodian.
For Australian agencies and research institutes, a practical design follows the Five Safes directly:
- Safe projects: connect project approval, ethics approval and data custodian sign-off in one auditable workflow.
- Safe people: tie workspace access to verified identity, training records and accreditation status under the DATA Act where relevant.
- Safe settings: deploy analysis environments inside the custodian’s own cloud or data center in Australia, with no internet egress from workspaces.
- Safe data: harmonize to shared standards such as the OMOP Common Data Model (CDM) and HL7 FHIR, and apply de-identification before researchers see records.
- Safe outputs: check every export for small cells, rare values and disclosive model artifacts before release.
Common pitfalls
Treating each dataset as a separate project
Researchers often wait months for approvals from several custodians, each with its own environment. Shared infrastructure that each custodian controls locally reduces duplication without forcing data into one place.
Building AI outside the secure setting
Exporting data to a separate compute platform for model training breaks the safe settings principle. Training should happen where the data is governed.
Assuming offshore cloud regions are equivalent
Many Australian government and health bodies expect data to be hosted in Australia by certified providers. Confirm hosting location and certification early.
Forgetting outputs from models
Model weights and generated text can disclose information about individuals. Treat them as outputs that need checking, just like tables.
What to do next
If you are planning a health data or AI program in Australia, start by mapping every data source to its custodian, legal basis and approval pathway: Privacy Act, state health privacy law, DATA Act, or a combination. Decide which environment each analysis will run in and whether that environment satisfies all Five Safes. Where data sits with several custodians, design for federation from the start so that no custodian has to give up control. Our explainer on what a Trusted Research Environment is is a useful starting point for stakeholders new to the model.
Frequently asked questions
How do researchers access health data in Australia?
Researchers usually apply to the relevant data custodian, obtain ethics approval from a human research ethics committee, and then analyze data inside a secure environment such as ABS DataLab, AIHW’s secure access or the Sax Institute’s SURE. Identifiable data is rarely released directly, and outputs are checked before release.
Does Australia use the Five Safes framework?
Yes. The ABS and the AIHW both use the Five Safes to govern access to sensitive data, and the data sharing principles in the Data Availability and Transparency Act 2022 are closely modeled on it: project, people, setting, data and output.
What is the Data Availability and Transparency Act 2022?
The DATA Act creates a national scheme for sharing Australian Government data with accredited users for permitted purposes, including research. It is overseen by the National Data Commissioner and relies on accredited users and accredited data service providers.
Which privacy law applies to health research in Australia?
The Privacy Act 1988 and its Australian Privacy Principles apply to Commonwealth agencies and many private organizations. State and territory laws apply to public hospitals and state health departments. Many projects must satisfy both, alongside NHMRC ethics guidance.
What is data linkage in Australia?
Data linkage joins records about the same person across datasets. Australian linkage, coordinated nationally through the Population Health Research Network, separates identifiers from clinical content so that no single party sees both.
How does a federated TRE support Sovereign AI in Australia?
A federated TRE keeps data with each custodian in Australian infrastructure and sends approved analysis and model training to it. Only reviewed results leave, so custodians keep control and AI can be trained on national data without moving it offshore or into a central store.