Diagnosis classification in EMR data using latent representations and SNOMED-CT mapping for improved medical data...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Ontology Mapping

Background:

• Integrating medical data is challenging due to heterogeneous electronic medical record (EMR) systems.
• Standardized clinical ontologies like Systematized Nomenclature of Medicine-Clinical Terms (SNOMED-CT) are crucial for data interoperability.
• Automated mapping of EMR diagnoses to SNOMED-CT is needed for efficient data integration.

Purpose of the Study:

• To develop a diagnosis classification model for mapping EMR diagnosis spans to SNOMED-CT.
• To leverage ClinicalBERT for creating latent representations of diagnosis spans.
• To investigate semantic alignment and structural characteristics of these representations.

Main Methods:

• Trained ClinicalBERT on EMR data to generate diagnosis span embeddings.
• Fine-tuned embeddings using mean squared error (MSE) to align with SNOMED-CT fully specified names (FSNs).
• Constructed a downstream classification model and analyzed latent space characteristics.

Main Results:

• Achieved high performance: 0.934 accuracy, 0.923 weighted F1-score, 0.823 macro-averaged F1-score on 273 SNOMED-CT classes.
• Demonstrated improved semantic alignment, reducing Manhattan and cosine distances by 36.2% and 53.1%, respectively.
• Showcased competitive performance against state-of-the-art models (SapBERT, BioSyn) with enhanced clustering coherence.

Conclusions:

• The proposed model effectively maps EMR diagnoses to SNOMED-CT, enhancing semantic understanding.
• This approach offers a scalable and privacy-preserving solution for medical concept normalization.
• Analysis provides insights into the balance between semantic alignment and classification performance.