Using Natural Language Processing to Extract and Classify Symptoms Among Patients with Thyroid Dysfunction.

Area of Science:

• Medical Informatics
• Clinical Natural Language Processing
• Endocrinology

Background:

• Thyroid dysfunction affects over 12% of the US population, often presenting with fatigue and weight changes.
• The relationship between these patient-reported outcomes and thyroid testing patterns remains unclear.
• Accurate identification of symptoms from clinical notes is crucial for understanding patient populations.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) pipeline for identifying fatigue and weight change symptoms in patients with thyroid stimulating hormone (TSH) tests.
• To compare the performance of various Bidirectional Encoder Representations from Transformers (BERT) models in extracting these specific symptoms from clinical text.

Main Methods:

• A natural language processing and deep learning pipeline was created to detect patient-reported outcomes of weight change and fatigue.
• Five open-source BERT models were compared for their accuracy in identifying these symptoms from clinical notes.
• Performance was evaluated using F1-scores for both fatigue and weight change detection.

Main Results:

• Bio_ClinicalBERT achieved the highest F1-scores for identifying fatigue (0.900) and weight change (0.906).
• Other evaluated models included BERT, DistilBERT, Biomedical RoBERTa, and PubMedBERT, with varying performance.
• The results indicate Bio_ClinicalBERT's superior capability in extracting these specific symptoms from clinical text.

Conclusions:

• The developed NLP pipeline, particularly using Bio_ClinicalBERT, can accurately identify fatigue and weight change in patients undergoing thyroid testing.
• This methodology can aid in analyzing patient-reported outcomes and understanding thyroid dysfunction prevalence and testing patterns across demographics.