OHSU Knight Federated Learning Network Prototype to Support Multimodal Cancer Models

Oregon Health & Science University, Portland OR

Investigators

Linked publications & trials

Abstract

PROJECT SUMMARY/ABSTRACT Traditional data-sharing models typically centralize de-identified data for analysis. Patient records have become more multifaceted and high-dimensional such that de-identification alone is inadequate to safeguard privacy. Privacy violations can result in harms to patients that can be both deontological (ethically problematic) or consequentialist (harm such as discrimination, financial consequences, and stigma). Current regulatory guidelines are not designed for the complexity of AI/ML. Moreover, each academic medical center will have their own ethics boards, governance models, privacy standards and security requirements. Risk of a privacy breach limits access and sharing of patient data — a barrier to research across cancer centers and academic medical centers. Given its outstanding strengths in precision oncology, computational biology, machine learning, clinical curation, data governance and data quality and regional data sharing, OHSU Knight Cancer Institute is uniquely situated to assist NCI in the development of a federated learning network. The long-term objective is to develop the best practices and guidance to support large-scale federated learning across cancer centers to achieve optimal benefit through broad data sharing. In the near term, the focus of this project is to support the development and testing of a federated learning network prototype. To do this, the OHSU Knight Cancer Institute proposes the following aims. Aim 1 is to develop a multi-modal deep learning model to predict tumor mutational burden in prostate cancer. Aim 2 is to implement and evaluate our federated learning prototype focusing on privacy, compliance and bias reduction. This will also include assessment of the impact of adversarial attacks and how well different defense strategies mitigate this. Finally, Aim 3 is the evaluation of the federated model on the federated learning network in collaboration with partner sites (other cancer centers, NCI).