Identifying symptom etiologies using syntactic patterns and large language models.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Differential diagnosis is essential for accurate medical practice.
• Current resources for identifying symptom causes are limited by manual curation.
• Novel methods are needed to extract comprehensive etiological information from scientific literature.

Purpose of the Study:

• To introduce and analyze two novel methods for mining symptom etiologies from scientific literature.
• To compare the coverage and precision of traditional NLP versus generative models for etiology extraction.
• To evaluate the synergistic benefits of combining both approaches.

Main Methods:

• A traditional Natural Language Processing (NLP) method using human-guided pattern bootstrapping for syntactic pattern extraction.
• A novel method employing generative models (GPT-4) with a fact verification pipeline.
• Comparative analysis of the coverage and precision of each method and their combined application.

Main Results:

• The NLP method achieved significant coverage in extracting symptom etiologies.
• The GPT-4 method demonstrated high precision but lower coverage compared to the NLP approach.
• Combining both methodologies resulted in enhanced depth and reliability of etiology mining.

Conclusions:

• Two novel methods for mining medical symptom etiologies from scientific literature have been presented.
• A hybrid approach combining NLP and generative AI offers superior depth and reliability for differential diagnosis support.
• These methods have the potential to improve the identification of novel or less common disease causes.