ROADMAP: An Ontology of Medical AI Models and Datasets

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Ontology Development

Background:

• Clear communication of AI models and datasets is crucial for medical AI development, regulatory review, and clinical use.
• Existing frameworks like "model cards" and "datasheets for datasets" lack comprehensive support for multimodal medical data.
• Need for a machine-interpretable framework to standardize the description of medical AI resources.

Purpose of the Study:

• To introduce the Radiology Ontology of AI Datasets, Models, and Projects (ROADMAP).
• To provide a machine-interpretable framework for formally defining attributes and values of AI models and datasets.
• To enhance discoverability, interoperability, and reuse of medical AI resources.

Main Methods:

• Developed ROADMAP as a formal ontology.
• Incorporated features for multimodal data (images, structured, unstructured text).
• Integrated concepts from existing ontologies, coding schemes, and common data elements.

Main Results:

• ROADMAP provides a machine-interpretable framework for describing medical AI resources.
• The ontology supports multimodal data types.
• It enhances discoverability, interoperability, and reuse of AI resources.
• Facilitates matching AI models with datasets and detecting potential bias.

Conclusions:

• ROADMAP offers a standardized, machine-interpretable approach to describing medical AI resources.
• The ontology promotes better data management, model development, and bias assessment in medical AI.
• ROADMAP is available at https://bioportal.bioontology.org/ontologies/ROADMAP.