The healthcare sector (public and private) generates an unparalleled amount of data from sources as diverse as electronic medical records, advanced imaging techniques, high-throughput sequencing, wearable devices, and public health data. Leveraging these massive datasets through sophisticated machine learning algorithms has the potential to transform medical practice by enabling the development of more effective and personalized treatments, interventions, and public policies, ultimately improving healthcare delivery and population well-being. However, the highly sensitive nature of health data, cybersecurity risks, biases in the data, and the lack of robustness in machine learning algorithms are key obstacles currently preventing the full realization of recent advancements in artificial intelligence.
To overcome these challenges, it is essential to address ethical, legal, security, and robustness issues. This project aims to develop new machine learning algorithms that account for the multi-scale and heterogeneous characteristics of health data while ensuring privacy, robustness against adversarial attacks and changes in data and model dynamics, and fairness for underrepresented populations. By addressing these obstacles, we hope to unlock the barriers that hinder the deployment of innovative solutions in digital health.
Specifically, the project will focus on the following challenges:
(i) privacy-preserving learning through differential privacy techniques and homomorphic encryption;
(ii) federated learning by balancing accuracy and privacy;
(iii) robustness against adversarial attacks and changes in data and model dynamics;
(iv) automated “forgetting” mechanisms to implement the right to be forgotten.
The project, part of Project of PEPR Digital Health (2023-2027), brings together a unique consortium of experts in machine learning, cybersecurity, statistics, and medical applications. Moreover, it is strategically positioned between two national programs (PEPRs: Cybersecurity and Digital Health), providing a unique opportunity to disseminate knowledge and best practices.
