Related Experiment Videos
Machine unlearning as a governance imperative for clinical AI
Anthony Porter1,2, Emily Kirkpatrick3, Arpit Garg3
1Australian Institute for Machine Learning (AIML), The University of Adelaide, Adelaide, SA, Australia. anthony.porter3@sa.gov.au.
NPJ Digital Medicine
|July 18, 2026
Summary
Clinical AI must be prepared for data removal. Machine unlearning enables removing specific data influence without full retraining, ensuring patient autonomy and clinical safety in healthcare AI.
Area of Science:
- Healthcare AI
- Machine Learning
- Data Governance
Background:
- Current clinical AI models assume data influence is permanent.
- This assumption is challenged by evolving evidence, patient consent withdrawal, and identified bias.
- The inability to remove data influence poses risks to patient autonomy and data integrity.
Purpose of the Study:
- To advocate for built-in "unlearning readiness" in high-risk healthcare AI infrastructure.
- To address the need for removing specific data influence without costly full retraining.
- To propose a governance framework for safe and auditable AI updates.
Main Methods:
- Conceptual framework development for unlearning readiness.
- Analysis of infrastructure requirements for patient autonomy, clinical validity, and system governance.
- Outline of a governance pathway for auditable and clinically safe AI updates.
Main Results:
- Unlearning readiness is crucial for adaptable and ethical clinical AI.
- Integrating unlearning capabilities enhances patient data control and trust.
- A structured governance pathway ensures safe deployment and maintenance of AI systems.
Conclusions:
- Healthcare AI infrastructure must proactively incorporate machine unlearning capabilities.
- This approach supports patient rights, maintains clinical accuracy, and ensures robust system governance.
- Implementing unlearning readiness is essential for the future of trustworthy and responsible clinical AI.