A corpus-driven standardization framework for encoding clinical problems with HL7 FHIR.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Data Standardization

Background:

• Free-text problem descriptions in medical records are difficult to standardize.
• Capturing the semantics of complex medical conditions is challenging.
• Lack of standardization hinders interoperability and data analysis.

Purpose of the Study:

• To develop a framework for transforming free-text problem descriptions into standardized Health Level 7 (HL7) Fast Healthcare Interoperability Resources (FHIR) models.
• To accurately extract clinical concepts and their relationships from free text.
• To facilitate the mapping of clinical problems to the FHIR Condition resource.

Main Methods:

• Utilized domain-specific dependency parsers, Bidirectional Encoder Representations from Transformers (BERT), and cui2vec Unified Medical Language System (UMLS) concept vectors.
• Employed a neural network classification model to identify thirteen relationship types between concepts.
• Applied data programming (weak supervision) and Shapley values for model interpretation.

Main Results:

• Achieved an F1 score of 0.95 for identifying the focus concept (primary clinical concern).
• Extracted relationships between concepts with an F1 score of 0.90.
• Obtained a 0.89 weighted average F1 score for relationship classification, enabling accurate mapping to FHIR models.
• Shapley values indicated BERT's significant contribution to classifier decisions.

Conclusions:

• The proposed framework effectively transforms free-text problem descriptions into structured HL7 FHIR models.
• The methods demonstrate high accuracy in concept and relationship extraction.
• This approach enhances the standardization and interoperability of clinical problem data.