A Clinical Reasoning-Encoded Case Library Developed through Natural Language Processing

Area of Science:

• Artificial Intelligence in Medicine
• Computational Linguistics
• Medical Informatics

Background:

• Clinical reasoning instruction relies on case reports, but their searchability by symptoms or diagnoses is limited.
• Developing computational methods to analyze diagnostic reasoning in case reports can enhance their educational utility.

Observation:

• A "reasoning-encoded" case database was developed using natural language processing (NLP) on 2525 New England Journal of Medicine (NEJM) Clinical Pathological Conference (CPC) cases (1965-2020).
• NLP identified 12 categories of medical terms and their relationships, extracting 43,291 symptoms and 6532 diagnoses.
• A measure of differential diagnosis similarity between cases was derived, and a website was created for exploring the data.

Findings:

• NLP algorithms successfully identified clinically relevant categories reflecting relationships between medical terms, including symptoms, signs, pathophysiology, and diagnoses.
• The analysis revealed patterns in how expert clinicians construct differential diagnoses.
• The developed database provides a novel way to explore diagnostic reasoning across a large case series.

Implications:

• This reasoning-encoded database offers insights into expert diagnostic processes, aiding clinicians in correlating disease categories.
• Clinician-educators can use this resource to design case-based curricula.
• Physicians can leverage the database for self-directed lifelong learning to enhance diagnostic skills.