Strategic Planning Workshop

March 11-12, 2025, 11AM – 5PM ET   |   Virtual

As part of strategic planning activities, the NIH Common Fund hosted a virtual workshop to engage subject matter experts, community members, and NIH staff in a discussion of high-priority areas for the NIH to consider when developing the PRIMED-AI program concept. The primary focus was to identify opportunities and challenges to enable precision medicine with artificial intelligence (AI) through the integration of clinical imaging and multimodal data. The session topics included:

1. Developing algorithms that leverage multimodal data to solve clinical needs
2. Accessing and preparing AI-ready imaging and multimodal data to enable interoperability
3. Validating and implementing clinical imaging and multimodal AI
4. Ethical considerations for the use of imaging based, multimodal AI clinical decision support tools

These topics were explored in a series of informational presentations and discussion sessions featuring various expert speakers and panelists. The discussion sessions were guided by questions from the panelists and from participants. An archived version of the agenda is available. 

Workshop Executive Summary

This summary represents the opinions and perspectives of the workshop participants, which do not necessarily reflect the perspectives of NIH, the federal government, or the goals or structure of the potential PRIMED-AI program.

On March 11-12, 2025, the National Institutes of Health (NIH) Office of Strategic Coordination –The Common Fund convened the Precision Medicine with AI: Integrating Imaging with Multimodal Data (PRIMED-AI) workshop to identify key opportunities, complexities, and challenges in the emerging space of AI for precision medicine. Experts across NIH, academia, and industry gathered to share perspectives in four sessions, summarized below.  The workshop aimed to chart a course for the integration of AI with imaging and other diverse health data types (multimodal data) to advance precision medicine.

Session 1, Developing Algorithms that Leverage Multimodal Data to Solve Clinical Needs, emphasized the crucial relationship between data and metadata for interpreting AI model-identified outputs. Multimodal health data includes both imaging data (e.g., CT, MRI, and others) and non-imaging data (e.g., electronic health records, medical reports, and others), spanning multiple formats and domains. Co-chairs and panelists discussed the roles that AI models can play in detecting and understanding biological processes as well as the role of collecting and preparing data to ensure that models can perform these functions accurately. Developing robust AI models often requires curation and harmonization of noisy data as well as training of the algorithm with high quality data, both of which are informed by metadata. Co-chairs and panelists also discussed ways in which AI may prove clinically beneficial, such as identifying disease subtypes. Key discussions centered on the need for well-annotated, high-quality multimodal data to build AI models capable of addressing specific clinical needs and uncovering valuable biological insights.

Session 2, Accessing and Preparing AI-Ready Imaging and Multimodal Data to Enable Interoperability, focused on the collection, curation, and provenance of datasets for use in AI development. AI model development requires large quantities of deidentified, harmonized, and labeled data. Co-chairs and panelists highlighted concerns and potential best practices around data sharing and accessibility to meet these needs and to maintain sustainable health databases by evaluating dataset representativeness and quality. The session underscored the critical importance of establishing robust mechanisms for accessing, preparing, and sharing large, high-quality, AI-ready datasets while ensuring data privacy, interoperability, and quality.

Session 3, Validating and Implementing Clinical Imaging and Multimodal AI, examined the AI implementation landscape, including the processes already in place to validate and monitor AI models in clinical settings and processes that could be developed in the future to support AI model implementation. Co-chairs and panelists considered essential issues including who (e.g., AI developers, federal agencies, independent governance groups, others) should play a role in AI model monitoring and what metrics should be used to evaluate AI model implementation beyond accuracy. Generalizability, fit-for-purpose, and clinical benefit are important metrics to consider when determining the value of the implemented AI model. Discussions in this session revolved around the crucial steps for validating and implementing AI models in clinical practice. 

Session 4, Ethical Considerations for the Use of Imaging-Based, Multimodal AI Clinical Decision Support Tools, highlighted patient privacy and concerns related to underlying assumptions for AI development and implementation as well as technical strategies and frameworks for accountability to mitigate these risks and ensure transparent, trustworthy AI development. The chair and panelists raised concerns about the relationships between patient consent and data collection for AI model development. This session addressed the significant ethical implications of using imaging-based, multimodal AI in clinical decision support. 

In conclusion, workshop participants emphasized the paramount importance of developing AI models that aim to provide clinical benefit and further patient health. Developing clinically useful AI models will essentially require representative health data across a variety of health conditions, disease states, and populations coupled with supportive data sharing practices and policies.  
 

A full report of the workshop can be viewed here.