US20260204427
2026-07-16
Physics
G16H50/70
The system processes unstructured clinical text by linking it to standardized medical codes using embedding-based similarity search. It accepts clinical language input via an API, which is then transformed into high-dimensional representations by a semantic embedding model. A hybrid search, combining vector similarity and keyword matching, is conducted against a database of medical terminology. This provides a ranked list of candidate codings and schemas, facilitating accurate mapping of clinical terms to standard vocabularies, essential for clinical analytics and decision support.
The healthcare sector generates vast amounts of unstructured data, such as electronic health records and clinical notes, which are rich in information but challenging to process due to language variability. Medical concepts can be expressed in numerous ways, complicating automated recognition. Standardized coding systems like SNOMED CT and ICD-10-CM offer structured vocabularies for clinical information. Effective normalization of unstructured text to these codes is crucial for leveraging clinical data in automated systems.
The disclosed techniques involve a computer-implemented method that processes unstructured text queries received via an API. A semantic embedding model generates a query vector representing the text's semantic meaning. A hybrid search, using this vector, is performed on a pre-indexed vector database containing medical codings and schemas. The search combines vector similarity with keyword searches to yield a ranked list of relevant codings and schemas, aiding in the creation of machine-readable representations of clinical queries.
Various embodiments include preprocessing text queries to extract key phrases, generating query vectors, and performing k-Nearest Neighbors (kNN) vector similarity searches. These searches are weighted and combined with keyword searches to produce relevant candidate codings. Structured responses are generated for downstream applications, and the system can execute actions based on these responses. The method also supports structured requests with pre-identified entities, using entity types to constrain searches for more precise results.
The system can handle structured requests with pre-identified text entities, generating query vectors for each entity and performing hybrid searches constrained by entity type. The search results are ranked by relevance and confidence scores, with the option to target specific coding systems. Named Entity Recognition (NER) processes can generate these entities, enhancing the system's ability to map text to standardized codes effectively and efficiently.