Key Takeaways
- A team of Vanderbilt University Medical Center researchers demonstrated AI tools like ChatGPT can effectively extract critical disease information from medical records, potentially accelerating rare disease diagnosis.
- The study shows AI-powered phenotyping could make rare disease analysis more accessible for healthcare facilities that lack resources for extensively trained AI systems, as they require minimal data preparation.
Researchers at Vanderbilt University Medical Center have released a study demonstrating that artificial intelligence tools could accelerate diagnosis and treatment for many patients with rare diseases.
The findings, reported in the Journal of Healthcare Informatics Research, showed that ChatGPT can match or outperform traditional approaches at identifying certain disease characteristics, even with minimal training data.
“This work demonstrates the potential of AI tools like ChatGPT to streamline rare disease diagnosis and accelerate patient care,” said Cathy Shyr, Ph.D., of the Department of Biomedical Informatics at Vanderbilt University Medical Center and lead researcher on this study.
A rare disease is defined as affecting fewer than 200,000 people individually. Yet collectively, rare diseases impact 300 million people worldwide. Such patients often face lengthy diagnostic journeys with devastating medical and personal consequences, making faster diagnosis a critical healthcare priority.
Phenotyping, analyzing a patient’s record to create a comprehensive picture of how a disease manifests, is critical for accelerating rare disease diagnosis. The data required to assemble a comprehensive picture of a patient’s history is often buried in unstructured data, making it challenging to automate the diagnostic process. Traditionally, highly trained experts manually comb through unstructured medical texts to identify critical disease characteristics.
Phenotyping with AI
Automatic disease recognition is particularly challenging given the diversity, complexity and specificity of rare diseases and their phenotypes. The study compared two different AI paradigms for extracting disease information using large language models: fine-tuning using BioClinicalBERT and prompt learning with ChatGPT.
“This work demonstrates the potential of AI tools like ChatGPT to streamline rare disease diagnosis and accelerate patient care.”
Fine-tuning requires extensive manual review and extraction of medical datasets. Prompt learning allows researchers to use a large language model, in this case ChatGPT, effectively “out of the box” with minimal preparation.
While BioClinicalBERT performed better overall, ChatGPT showed surprising strengths with minimal training data, achieving higher accuracy than the more labor-intensive approach at identifying rare diseases (77.8 percent) and clinical signs (72.5 percent).
Achieving comparable results without costly, highly skilled labor has the potential to make this AI model accessible for healthcare facilities that don’t have resources for extensively trained AI systems.
Moving towards Clinical Practice
The team is working on ways to implement these AI tools at the Vanderbilt Undiagnosed Diseases Network.
“Patients referred to the Undiagnosed Diseases Network have often spent years on a diagnostic odyssey with extensive workups, but no clear answers,” Shyr said.
Along with shortening the diagnostic journey, the AI method can identify patients suitable for clinical trials, which could accelerate the development of new treatments for these patients.
Looking Ahead
While this research is promising, Shyr emphasized that AI tools like these are meant to complement, not replace, medical expertise.
“The goal is never to replace the physician or clinician, but to leverage AI tools to streamline the process.”
“The goal isn’t to replace clinicians, but to leverage AI tools to streamline the diagnostic process. The question is: how do we reduce the time it takes to reach a diagnosis? That’s where these models can truly shine.”
