06-06-2023 дата публикации
Номер: CN116227433A
Принадлежит:
The invention relates to a medical knowledge injection prompt-based few-sample ICD coding method and system, and the method can generate an optimal ICD code according to an input medical text, and improves the matching accuracy between the medical text and the ICD code. The method comprises the following steps: S1, preprocessing an input medical text; s2, constructing a data set through synonyms, abbreviations and hierarchical structures in a medical knowledge graph of UMLS and ICD ontologies, pre-training a longform model by using hierarchical triplet loss, and injecting structured medical domain knowledge into the longform model; and S3, generating a corresponding code description for each ICD code c by using the UMLS, splicing the preprocessed medical text t, the code description and the fixed text template together to form a new input sequence, and classifying the input sequence by using the trained longformer model to obtain a classification result of the ICD codes.
Подробнее