Mitigation of privacy issues in the development of clinical AI models

Sep 19, 2024, 1:00 PM
15m
Adam-von-Trott-Saal

Adam-von-Trott-Saal

Session 5. AI-Applications in Clinical Practice: Challenges and Successes 🇬🇧 Session 5. AI-Applications in Clinical Practice: Challenges and Successes

Speaker

Stefan Schrod (University Medial Center Göttingen)

Description

The development of highly accurate AI models for clinical applications, such as disease classification or personalized treatment recommendation, is heavily constrained by small amount of available training data. The data is commonly generated at different institutions, subject to different data generating processes and privacy regulations. In order to maximize the predictive performance, all of these challenges have to be addressed simultaneously.

Commonly used Federated Learning (FL) approaches solely aggregate the knowledge of multiple confidential data source by distributing the model development. However, they do not address the high heterogeneity of the different data sources, i.e., due to different experimental equipment, which can introduce significant bias for unseen data.

To address this issue we introduce Federated Adversarial Cross Training (FACT). FACT uses the implicit domain differences between the clients to identify domain shifts. In each round of FL, FACT cross initializes a pair of clients to generate domain specialized representations which are subsequently used as a direct adversary to learn a domain invariant data representation. We show that FACT not only outperforms state-of-the-art federated, non-federated and source-free domain adaptation models on popular FL benchmarks, but also improves AI model development in clinical applications.

Primary authors

Stefan Schrod (University Medial Center Göttingen) Jonas Lippl (University Medial Center Göttingen) Michael Altenbuchinger (University Medial Center Göttingen)

Presentation materials

There are no materials yet.