Blog
Can AI Learn from Health Data Without Exposing Patient Information?
Artificial intelligence has the potential to transform healthcare, from supporting earlier disease detection to helping clinicians make better-informed decisions. Yet many of these applications depend on access to sensitive health data, raising an important question: how can we unlock the value of medical data without compromising patient privacy?
Healthcare organisations face a difficult balance. Sharing data can support medical research, improve clinical decision-making and enable more personalised care. At the same time, health records contain deeply personal information that must be protected. Finding ways to use this data responsibly is becoming increasingly important as AI becomes more integrated into healthcare systems.
The challenge: Making health data useful while keeping it private
AI systems can identify patterns across large volumes of information, potentially helping healthcare professionals recognise risks and make more informed decisions. However, using sensitive medical information introduces risks that cannot be ignored. Unauthorised access, excessive data sharing and inadequate safeguards can undermine patient trust and expose individuals to harm.
Traditional approaches often rely on transferring data to a central location for processing. While this can simplify data analysis, it may also create additional privacy and security challenges, particularly when multiple organisations need to collaborate. Removing names or other direct identifiers is not always sufficient, as individuals may sometimes be re-identified by combining different pieces of information.
The challenge is therefore not simply to anonymise health records. It is to develop approaches that allow useful computation while limiting unnecessary exposure of sensitive information.
Privacy-enhancing technologies: A different approach
Privacy-enhancing technologies (PETs) offer ways to reduce the amount of sensitive information that needs to be exposed during data processing. Rather than relying exclusively on restricting access to data, these technologies can help protect information while it is being used.
Different approaches address different challenges. Federated learning, for example, can allow AI models to be trained across multiple locations without routinely centralising all the underlying data. Secure computation techniques can help parties perform calculations while limiting what they learn about one another’s inputs. Privacy-preserving inference focuses on protecting data involved in the process of generating predictions from an AI model.
These approaches are not interchangeable, and none automatically eliminates every privacy risk. Their effectiveness depends on the system design, implementation, threat model and operational context. Nevertheless, they offer important opportunities to make data-driven innovation more compatible with privacy requirements.
From technology to practice: The LICORICE eHealth pilot
The Horizon Europe project LICORICE is exploring how privacy-preserving technologies can support trusted digital interactions and sensitive data processing. Its eHealth pilot brings together two important aspects of this challenge: giving individuals greater control over their data and enabling clinicians to use AI-supported risk assessment.
From the patient’s perspective, the pilot explores the use of the uSelf Self-Sovereign Identity (SSI) wallet for authentication and the management of data-sharing permissions. The aim is to support a workflow in which patients can manage how their information is shared, rather than treating consent and identity verification as invisible technical steps.
From the clinician’s perspective, the pilot explores privacy-preserving neural network inference (pp-NNI) for AI-supported risk assessment. This approach focuses on protecting sensitive information during the processing needed to generate AI predictions. The workflow also considers explainable AI (XAI), helping clinicians interpret the resulting risk predictions as part of the decision-making process.
Together, these elements address complementary questions: how can patients exercise control over their information, and how can healthcare professionals access useful AI-supported insights while respecting privacy requirements?
The project also recognises that technical protection alone is not enough. Privacy controls need to be understandable, system feedback needs to be clear, and workflows must fit the needs of the people using them. This is why usability evaluation and stakeholder feedback are important parts of developing solutions for real-world healthcare settings.
Building trust in AI-enabled healthcare
The future of healthcare AI will depend not only on the accuracy of its predictions, but also on whether patients, clinicians and healthcare organisations can trust how data is handled. Privacy-preserving technologies can contribute to this trust by reducing unnecessary exposure and supporting more controlled processing, but they must be combined with appropriate governance, security measures and meaningful human oversight.
LICORICE contributes to this effort by exploring how privacy-preserving inference, digital identity and data-sharing controls can work together in an eHealth setting. Its work reflects a broader principle: protecting patient privacy and enabling healthcare innovation should be pursued together, not treated as opposing goals.
AI may offer significant opportunities for better healthcare, but those opportunities must be developed responsibly. By designing systems that consider privacy, usability and clinical needs from the outset, projects such as LICORICE are helping explore how sensitive health data can support innovation while keeping patient trust at the centre.
