Annotation and Information Extraction of Consumer-Friendly Health Articles for Enhancing Laboratory Test Reporting

Area of Science:

• Natural Language Processing
• Medical Informatics
• Patient Education

Background:

• Patient portal use is high, with lab result viewing being the most frequent activity.
• Lab results are often confusing for patients, hindering comprehension and engagement.
• Existing research offers limited context-specific interpretations for individual patient data.

Purpose of the Study:

• To evaluate transformer-based language models for recognizing and classifying key terms in laboratory test results.
• To map identified terms to controlled terminologies for enhanced data standardization.
• To lay the groundwork for improved, contextualized lab result presentations in patient portals.

Main Methods:

• Collected and annotated interpretations from 251 health articles on laboratory tests.
• Evaluated BioBERT, ClinicalBERT, RoBERTa, and PubMedBERT for term recognition and typing.
• Utilized BioPortal's term search API to map annotated terms to SNOMED CT, LOINC, and ICD-10-CM.
• Assessed model performance using strict and lenient matching F1 scores.

Main Results:

• PubMedBERT demonstrated the highest performance (F1 score) in recognizing key terms.
• SNOMED CT provided the best coverage for mapping annotated terms, followed by LOINC and ICD-10-CM.
• The study successfully identified and mapped critical terms within health articles related to lab tests.

Conclusions:

• Transformer models, particularly PubMedBERT, show significant promise for processing clinical text.
• Controlled terminologies like SNOMED CT are crucial for standardizing and understanding medical terms.
• This research enables the development of patient-centric tools for better interpretation of laboratory results.