Large language model-based scribing tools for ophthalmology: performance and safety evaluation using simulated...

Area of Science:

• Ophthalmology
• Medical Informatics
• Artificial Intelligence

Background:

• Six large language model (LLM)-based scribing tools were evaluated for their ability to generate ophthalmological clinical notes.
• The study focused on performance and safety in simulated physician-patient encounters.

Purpose of the Study:

• To compare the performance and safety of six LLM-based scribing tools in generating ophthalmological clinical notes.
• To identify the best-performing tool and assess the types and frequency of errors made by these AI scribes.

Main Methods:

• A cross-sectional comparative study used seven ophthalmic encounters transcribed by six scribing tools.
• Note quality was assessed by consultant ophthalmologists using the Physician Documentation Quality Instrument (PDQI-9) and text-generation metrics (ROUGE-L, BERTScore, BARTScore, AlignScore).
• Safety analysis identified documentation errors, including omissions, incorrect information, and extraneous content.

Main Results:

• Heidi AI achieved the highest overall documentation quality (PDQI-9: 38.524) and led in accuracy, usefulness, organization, comprehensiveness, and succinctness.
• Heidi AI also led in ROUGE-L (0.326), BERTScore (0.727), and AlignScore (0.704).
• All tools exhibited errors, primarily omissions of critical clinical details, but also incorrect or extraneous information.

Conclusions:

• Heidi AI demonstrated the highest performance among the evaluated LLM scribing tools for ophthalmology documentation.
• Despite promising performance, all tools exhibited errors, particularly omissions of critical details, posing potential risks to patient care.
• Human oversight and rigorous verification are essential before clinical integration of these AI scribing tools.