Meta-Predicates for Trustworthy Clinical AI Decision Support
A novel framework has been developed that incorporates meta-predicates to impose epistemological constraints within clinical decision support systems, meeting the regulatory standards for auditability as outlined by the EU AI Act and FDA guidelines. This methodology categorizes annotations across four key dimensions: purpose, knowledge domain, scale, and method of acquisition. It is implemented in AnFiSA, an open-source platform dedicated to the curation of genetic variants, and its effectiveness is showcased through the Brigham Geno dataset.
Key facts
- Regulatory frameworks like EU AI Act and FDA guidance require auditability for AI/ML-based medical devices.
- Existing formal languages validate syntactic and structural correctness but not epistemological appropriateness.
- Meta-predicates are predicates about predicates that assert epistemological constraints on clinical decision rules.
- The epistemological type system classifies annotations along four dimensions: purpose, knowledge domain, scale, and method of acquisition.
- The framework is instantiated in AnFiSA, an open-source platform for genetic variant curation.
- Demonstrated using the Brigham Geno dataset.
- The approach draws on design-by-contract principles.
- Meta-predicates specify which evidence types are permissible in any given rule.
Entities
Institutions
- EU AI Act
- FDA
- AnFiSA