Hepatology e-consult responses generated by artificial intelligence demonstrate accuracy but require human oversight

Area of Science:

• Hepatology
• Artificial Intelligence
• Medical Informatics

Background:

• Electronic consultations (e-consults) enhance specialist access but increase provider workload.
• A novel customized large language model (LLM), LiVersa, was developed for liver disease e-consults.
• The study evaluates LiVersa's efficacy in drafting responses and compares human and machine reviewer assessments.

Purpose of the Study:

• To assess the performance of LiVersa in generating hepatology e-consult responses.
• To evaluate the equivalence between human expert reviewers and an LLM-as-a-judge system.
• To determine the accuracy, safety, and usability of LLM-generated e-consult drafts.

Main Methods:

• LiVersa-generated responses for 61 hepatology e-consults from UCSF (Jan-Mar 2025) were analyzed.
• Three hepatologists and an "LLM-as-a-judge" system evaluated drafts using a 12-item rubric.
• Equivalence testing between human and LLM reviewers was performed using two one-sided tests (TOST).

Main Results:

• LiVersa drafts matched original responses in word count and verbosity.
• Human reviewers found 72% of drafts reasonable starting points and 83% provided appropriate recommendations.
• LLM-as-a-judge was more stringent, identifying more potentially harmful drafts (67% vs. 20%) and fewer clinically equivalent responses (27% vs. 48%) than human reviewers.

Conclusions:

• Customized LLMs like LiVersa demonstrate potential for streamlining e-consult drafting.
• Human oversight remains essential for ensuring the safety and accuracy of LLM-generated responses.
• LLM-as-a-judge serves as a conservative tool for quality assurance, particularly during LLM updates.