Clinical and Technical AI Safety Assessment
DHSCA is preparing to offer Clinical and Technical Assessment of AI-enabled tools, using the GUARDS™ Methodology. We are in the process of engaging an Australian University to externally validate GUARDS before commencing assessments.
GUARDS™ methodology - Clinical and Technical Assessment for AI-enabled tools
GUARDS™ is a proprietary, structured framework developed by DHSCA to evaluate AI-enabled tools deployed in clinical settings.
The framework applies to AI-enabled tools that are not TGA-regulated and sit below the Software as a Medical Device threshold. It addresses a significant governance gap: AI tools are being adopted across Australasian health and care systems at pace, without independent, standardised evaluation.
Every GUARDS™ assessment produces an AI Evaluation Report (AER) issued by DHSCA, following a five-stage pathway. The AER records domain scores, findings and one of three outcomes: Recommended for Deployment, Conditionally Recommended, or Not Recommended.
AERs are valid for 24 months, with 7 material-change triggers mandating re-assessment regardless of validity.
Read more about the GUARDS domains below
Governance and Accountability
Governance and Accountability assesses whether clear, enforceable lines of accountability exist for the AI tool’s clinical performance and whether governance arrangements are person-centred.
Person-centredness in this context means governance structures that explicitly place patient safety, dignity, and autonomy at the centre of decision-making and practice; not merely as a principle but as a measurable, monitored commitment embedded in the organisation’s controls and strategic intent.
Risk, Safety
and People
This domain assesses whether the tool can fail safely and whether the organisation has identified, managed and mitigated the clinical and human risks of deployment.
The “People” dimension ensures that workforce impacts, cultural safety, patient vulnerability, clinician burden and equity considerations are explicitly evaluated alongside technical risk.
Utility and Clinical Validation
Utility and Clinical Validation assesses whether the tool demonstrably does what it claims, in the population it will be used with.
Training data provenance and bias is assessed here, under algorithmic utility. This concerns whether the data used to train the model is representative, documented and assessed for bias in the affected population.
Data Governance and Privacy
Data Governance and Privacy assesses the data environment in which the tool operates. This domain is distinct from Utility (Domain U) which addresses training data and algorithmic utility.
Domain D addresses how patient and operational data is collected, stored, accessed, shared and protected during deployment.
Algorithmic Integrity
Algorithmic Integrity assesses whether the underlying model is documented, understood and traceable.
This domain is particularly critical for tools used in clinical decision support and risk stratification, where the interpretability of model outputs directly affects clinician trust and patient safety.
Surveillance and Oversight
Surveillance and Oversight assesses whether the tool will continue to be monitored after deployment and whether the organisation has the mechanisms to act on what monitoring reveals.
GUARDS treats post-market surveillance as a non-negotiable requirement, with assessment triggers built to manage mandatory re-assessment