Use of large language models to optimize poison center charting

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Clinical documentation

Background:

• Efficient medical charting is crucial for patient care and research.
• Current charting methods can be time-consuming and labor-intensive.
• The integration of AI offers potential solutions for improving documentation efficiency.

Purpose of the Study:

• To evaluate the capability of Chat Generative Pre-Trained Transformer (GPT) to generate medical charts from real-world poison center calls.
• To assess the accuracy and suitability of AI-generated charts for medical records.
• To compare AI-generated charts with traditional documentation methods.

Main Methods:

• De-identified transcripts of poison center calls were processed by Chat GPT 4.0.
• AI summarized calls and organized data into tables (vital signs, test results, therapies, recommendations).
• Seven medical experts reviewed and graded the AI-generated summaries for appropriateness.

Main Results:

• Eighty percent of AI-generated summaries met criteria for medical record entry.
• 91% of data points were accurately abstracted into tables.
• Reviewers preferred AI-generated charts, even those with initial errors.

Conclusions:

• Large language models can efficiently generate coherent medical summaries from audio recordings.
• AI-generated charts present a significant opportunity for improving efficiency in medical documentation.
• Future work will focus on prospective implementation and refinement of AI in clinical settings.