TL;DR
This paper explores the critical need for robust governance frameworks in biomedical data repositories to ensure the responsible development and deployment of AI/ML applications. It highlights the ethical, legal, and practical challenges encountered by four distinct NIH-funded projects, offering valuable lessons for future data initiatives.
What Problem Does This Paper Address?
The rapid advancement of AI and ML in healthcare promises transformative breakthroughs, but its effective and ethical implementation hinges on access to high-quality, well-governed data. Biomedical data repositories, the bedrock of this progress, face significant challenges in ensuring that data is collected, stored, shared, and used responsibly. These challenges span from defining what data to collect and how to obtain informed consent, to establishing secure storage, managing data access, and controlling data usage. Without a comprehensive governance strategy, there's a risk of perpetuating biases, compromising patient privacy, and hindering the equitable advancement of AI in medicine. This paper addresses this gap by examining the multifaceted governance issues faced by four large-scale biomedical data generation projects funded by the NIH's Bridge2AI program.
Methods
The paper outlines a framework for data governance, detailing six core steps involved in data sourcing and its use for AI model development:
- Determine and justify what data are selected: This involves assessing the rationale behind data collection, considering factors like resource availability, data quality standards (e.g., FHIR, OMOP CDM, GA4GH), and comparability.
- The role of consent varies depending on data source: The paper differentiates between studies requiring direct patient interaction (prospective informed consent) and those using existing data (retrospective, often exempt from consent but requiring strict compliance with laws).
- Determine where and how data are stored: This section discusses the trade-offs between on-premises and cloud-based storage, as well as emerging strategies like federated learning and trusted research environments.
- Determine what control participants have over data sharing: This addresses the balance between broad data sharing for research advancement and individual privacy concerns, including the role of Certificates of Confidentiality (CoCs).
- Data access and allowable uses: This involves defining who makes access decisions, who can access data, for what purposes, and under what conditions.
- Determine under what conditions data and models can leave the repository: This focuses on the challenges of enforcing data use agreements and preventing misuse, especially when data or models are exported.
The authors also emphasize the cross-cutting importance of "increasing attention to public concerns" throughout all these steps, advocating for community input and stakeholder engagement.
Datasets
The paper discusses four "new flagship biomedical and behavioral datasets" funded by the NIH's Bridge2AI Program:
- Precision Public Health (Voice): Collects voice data linked to health conditions.
- Cell Maps for AI (CM4AI): Focuses on spatiotemporal data from canonical cell lines.
- Program in Clinical Care AI (CHORUS Network): Gathers multimodal data to improve recovery from acute and critical illness.
- Artificial Intelligence Ready and Equitable Atlas for Diabetes Insights (AI-READI): Develops a multimodal dataset, particularly focusing on retinal imaging, for type 2 diabetes mellitus research.
These datasets are designed to be FAIR (findable, accessible, interoperable, and reusable) and are intended for use in future AI model development, rather than being tied to specific initial hypotheses.
Key Insights and Ideas
- "Ethics by Design" is paramount: Responsible data collection and governance must be considered from the very inception of a project, not as an afterthought.
- Consent is complex and context-dependent: The necessity and form of participant consent vary significantly based on the data source and how it was collected, requiring careful navigation of legal and ethical landscapes.
- Public engagement is crucial for trust: Actively involving the public, including patient advocacy groups and diverse communities, in data governance discussions is essential for building trust and ensuring that data use aligns with societal values.
- Data storage and access require careful consideration: Balancing the need for accessibility with robust security measures, and defining clear access protocols and allowable uses, are critical for protecting sensitive data.
- Enforcement of data use agreements is challenging: While data use agreements are standard, their enforceability and the practicalities of monitoring compliance present ongoing challenges.
- Federated learning offers a promising privacy-preserving approach: This technique allows AI models to be trained without directly sharing raw data, mitigating some privacy risks.
Key Results
The paper doesn't present quantitative results in terms of AI model performance but rather focuses on the qualitative challenges and strategies employed by the Bridge2AI projects. Key findings include:
- Diversity of approaches: The four DGPs have adopted varied strategies for data collection, storage, and access, reflecting the unique nature of their datasets and research goals.
- Incorporation of public input: Projects have actively sought public input through questionnaires, stakeholder forums, focus groups, and citizen science programs to address concerns about accuracy, fairness, nondiscrimination, and privacy.
- Varied data storage solutions: Data are stored in different environments, ranging from publicly accessible repositories (CM4AI) to controlled-access platforms requiring data licenses (AI-READI) and multi-enclave approaches (CHORUS).
- Defined access criteria: Access to data is managed through mechanisms like Data Access Committees and data use agreements, with varying levels of stringency depending on data sensitivity.
Limitations and Future Work
- Enforceability of data use agreements: The paper acknowledges the difficulty in ensuring that data use agreements are fully enforceable and that misuse is adequately deterred.
- Sustainability of repositories: The long-term sustainability of data repositories and the funding required for ongoing monitoring and stewardship remain significant concerns.
- Balancing broad access with privacy: The inherent tension between the NIH's encouragement of broad data sharing and the need to protect individual privacy is a persistent challenge.
- Future work: The authors suggest that future programs can learn from the Bridge2AI strategies. They also highlight the need for more effort in identifying and addressing the interests of patients and research participants, and for sponsors to provide resources for increased public engagement in data governance. Assessing the value of deliberative engagement and including participants on bodies charged with promoting responsible data collection and use are also recommended.
Summary updated