Technologies
This technology is a neural-symbolic AI system that translates ambiguous clinical text and abbreviations into standardized medical codes by combining language model reasoning with deterministic terminology matching to enable accurate data standardization.
Clinical records and biomedical literature are full of ambiguous abbreviations and non-standard terms. For example, a shorthand like “RA” can mean rheumatoid arthritis, right atrium, or renal artery depending on context. Linking these free texts to standard codes is foundational for EHR interoperability, clinical trial matching, and health analytics. However, existing tools force a tradeoff between rule-based systems that handle common terms but break on abbreviations and rare concepts, while AI models handle ambiguity better but fail unpredictably on uncommon terminology and are not interpretable. Healthcare organizations need a solution that handles the full range of clinical languages while remaining accurate, scalable, and transparent enough for regulated clinical workflows.
This technology is a neural-symbolic AI system called Medical Concept Mapping (MCM) that links free-text clinical mentions to standard terminology codes. Unlike a static pipeline that processes every term the same way, MCM operates as an AI agent. Specifically, it evaluates each incoming term, selects the appropriate processing path, executes it, and self-verifies the results before returning a code. When a term is clear, MCM uses fast, deterministic lookup against standard vocabularies. When a term is ambiguous, MCM calls a large language model to interpret the term in context, verify the interpretation preserves the original meaning, and then map the clarified term to the correct code. This selective, decision-driven approach keeps the system efficient while maintaining a full, auditable record of every step.
In benchmarking across three public datasets, MCM outperformed leading tools by up to 24.8 points on ambiguous abbreviations, with independent reviewers validating 79.4% of AI-generated interpretations as correct.
Chunhua Weng, Ph.D, FACMI, FIASHI