The medical field is buzzing with the potential of large language models (LLMs). These AI tools can handle various tasks, including summarizing text. As these tools start to pop up in hospitals, it's crucial to check how well they work. A study looked at 100 random adult visits to an emergency department (ED) between 2012 and 2023. The focus was on how well GPT-4 and GPT-3. 5-turbo could create summaries of these visits. The study checked for three main issues: inaccurate information, made-up information, and missing important details.

The results showed that 33% of summaries by GPT-4 and 10% by GPT-3. 5-turbo were completely error-free. GPT-4 did well with accuracy, but 42% of its summaries had made-up information. Meanwhile, 47% of the summaries left out key clinical details. Most inaccuracies and made-up info were found in the "Plan" section of the summaries. Omissions were mainly in the "Physical Examination" and "History of Presenting Complaint" sections.

The study also looked at how harmful these errors could be. On a scale of 1 to 7, the average harm score was 0. 57. Only three errors scored 4 or higher, which means they could cause lasting harm. This shows that while LLMs can create accurate summaries, they often miss or make up important details. Understanding where and how these errors happen is vital. This knowledge helps doctors review AI-generated content and avoid harming patients. It's clear that AI tools need careful oversight in medical settings.

The study also highlighted the need for better training and guidelines for using AI in healthcare. Doctors should be aware of the limitations of these tools. They must know how to spot and correct errors in AI-generated summaries. This way, they can ensure patient safety and the quality of care. It's also important to remember that AI is just a tool. It should assist doctors, not replace them. The final responsibility for patient care always lies with the healthcare professionals.

How Reliable Are AI Tools in Emergency Rooms?

Actions