Disclosure control of machine learning models from trusted research environments (TRE): New challenges and opportunities

Leonhard Med, a trusted research environment for processing sensitive research data

This paper provides an overview of the development and operation of the Leonhard Med Trusted Research Environment (TRE) at ETH Zurich. Leonhard Med gives scientific researchers the ability to securely work on sensitive research data. We give an overview of the user perspective, the legal framework for processing sensitive data, design history, current status, and operations. Leonhard Med is an efficient, highly secure Trusted Research Environment for data processing, hosted at ETH Zurich and operated by the Scientific IT Services (SIS) of ETH. It provides a full stack of security controls that allow researchers to store, access, manage, and process sensitive data according to Swiss legislation and ETH Zurich Data Protection policies. In addition, Leonhard Med fulfills the BioMedIT Information Security Policies and is compatible with international data protection laws and therefore can be utilized within the scope of national and international collaboration research projects. Initially designed as a "bare-metal" High-Performance Computing (HPC) platform to achieve maximum performance, Leonhard Med was later re-designed as a virtualized, private cloud platform to offer more flexibility to its customers. Sensitive data can be analyzed in secure, segregated spaces called tenants. Technical and Organizational Measures (TOMs) are in place to assure the confidentiality, integrity, and availability of sensitive data. At the same time, Leonhard Med ensures broad access to cutting-edge research software, especially for the analysis of human -omics data and other personalized health applications.

Efficacy of federated learning on genomic data: a study on the UK Biobank and the 1000 Genomes Project

Combining training data from multiple sources increases sample size and reduces confounding, leading to more accurate and less biased machine learning models. In healthcare, however, direct pooling of data is often not allowed by data custodians who are accountable for minimizing the exposure of sensitive information. Federated learning offers a promising solution to this problem by training a model in a decentralized manner thus reducing the risks of data leakage. Although there is increasing utilization of federated learning on clinical data, its efficacy on individual-level genomic data has not been studied. This study lays the groundwork for the adoption of federated learning for genomic data by investigating its applicability in two scenarios: phenotype prediction on the UK Biobank data and ancestry prediction on the 1000 Genomes Project data. We show that federated models trained on data split into independent nodes achieve performance close to centralized models, even in the presence of significant inter-node heterogeneity. Additionally, we investigate how federated model accuracy is affected by communication frequency and suggest approaches to reduce computational complexity or communication costs.

Are the European reference networks for rare diseases ready to embrace machine learning? A mixed-methods study

BACKGROUND

The delay in diagnosis for rare disease (RD) patients is often longer than for patients with common diseases. Machine learning (ML) technologies have the potential to speed up and increase the precision of diagnosis in this population group. We aim to explore the expectations and experiences of the members of the European Reference Networks (ERNs) for RDs with those technologies and their potential for application.

METHODS

We used a mixed-methods approach with an online survey followed by a focus group discussion. Our study targeted primarily medical professionals but also other individuals affiliated with any of the 24 ERNs.

RESULTS

The online survey yielded 423 responses from ERN members. Participants reported a limited degree of knowledge of and experience with ML technologies. They considered improved diagnostic accuracy the most important potential benefit, closely followed by the synthesis of clinical information, and indicated the lack of training in these new technologies, which hinders adoption and implementation in routine care. Most respondents supported the option that ML should be an optional but recommended part of the diagnostic process for RDs. Most ERN members saw the use of ML limited to specialised units only in the next 5 years, where those technologies should be funded by public sources. Focus group discussions concluded that the potential of ML technologies is substantial and confirmed that the technologies will have an important impact on healthcare and RDs in particular. As ML technologies are not the core competency of health care professionals, participants deemed a close collaboration with developers necessary to ensure that results are valid and reliable. However, based on our results, we call for more research to understand other stakeholders' opinions and expectations, including the views of patient organisations.

CONCLUSIONS

We found enthusiasm to implement and apply ML technologies, especially diagnostic tools in the field of RDs, despite the perceived lack of experience. Early dialogue and collaboration between health care professionals, developers, industry, policymakers, and patient associations seem to be crucial to building trust, improving performance, and ultimately increasing the willingness to accept diagnostics based on ML technologies.

Machine learning models in trusted research environments - understanding operational risks

Introduction

Trusted research environments (TREs) provide secure access to very sensitive data for research. All TREs operate manual checks on outputs to ensure there is no residual disclosure risk. Machine learning (ML) models require very large amount of data; if this data is personal, the TRE is a well-established data management solution. However, ML models present novel disclosure risks, in both type and scale.

Objectives

As part of a series on ML disclosure risk in TREs, this article is intended to introduce TRE managers to the conceptual problems and work being done to address them.

Methods

We demonstrate how ML models present a qualitatively different type of disclosure risk, compared to traditional statistical outputs. These arise from both the nature and the scale of ML modelling.

Results

We show that there are a large number of unresolved issues, although there is progress in many areas. We show where areas of uncertainty remain, as well as remedial responses available to TREs.

Conclusions

At this stage, disclosure checking of ML models is very much a specialist activity. However, TRE managers need a basic awareness of the potential risk in ML models to enable them to make sensible decisions on using TREs for ML model development.