Making head and neck cancer clinical data Findable-Accessible-Interoperable-Reusable to support multi-institutional...

Area of Science:

• Artificial Intelligence
• Bioinformatics
• Data Science

Background:

• Federated learning (FL) offers a privacy-preserving approach for AI development using distributed clinical data.
• A key challenge in FL is ensuring data is Findable, Accessible, Interoperable, and Reusable (FAIR).
• Current methods often require a common data schema, hindering collaboration across diverse institutions.

Purpose of the Study:

• To demonstrate that FAIR data principles facilitate collaboration in privacy-aware data exploration, visualization, and model training within FL consortia.
• To introduce a flexible method for making distributed clinical data FAIR without altering original data structures.

Main Methods:

• A "Schema-on-Read" FAIR-ification strategy was developed, involving decoupling data content from schema, semantic ontology annotation, and readout via semantic queries.
• Open-source tools were provided as Docker containers for on-premises data preparation by local investigators.

Main Results:

• A federated, privacy-preserving visualization dashboard was created for exploring 5 distributed datasets lacking a common origin schema.
• Robust and flexible prognostication models were developed and validated by integrating diverse data sources, including clinical risk factors and radiomics.

Conclusions:

• The proposed FAIR-ification procedure enables successful data reuse in FL consortia without mandating a common schema at data origin.
• This work supports the broader adoption of FL in healthcare AI by providing practical methods for enhancing data FAIRness.