By Melinda Young
Researchers found success using artificial intelligence (AI) to create a new care transition product — an easy-to-read discharge summary. The new product took the writing to a 6th-grade level from an 11th-grade level and scored higher on a key assessment tool.1
“Human-written discharge summaries are often written with technical language, abbreviations, and acronyms. That’s one reason why they’re [found to be] at a higher grade level,” says Jonah Zaretsky, MD, lead study author and a hospitalist, clinical assistant professor, and associate chief of medicine at NYU Langone Hospital Brooklyn. “We wanted to improve the understanding of patients when they were discharged from the hospital. It’s a critical time, and patients are getting lots of information. We wanted to make sure they had information in a way they could read, understand, and retain.”
Zaretsky and colleagues used Microsoft Azure OpenAI to turn 50 provider-created discharge summaries into a patient-friendly format. The original discharge summaries were largely written by people but had some autopopulated sections, such as lab results, Zaretsky notes.
Using prompt engineering, Zaretsky and colleagues considered what makes reading material understandable from a patient’s perspective and had the AI tool create discharge summaries within those parameters. They based it on society guidelines and principles of understanding in patient materials.
“We tried to give them the active voice, tried to make sure information was chunked into smaller, digestible paragraphs, had headings in paragraphs that were bolded to distinguish paragraphs, kept sentences relatively short, and used 6th- to 7th-grade reading levels, which is what is recommended for patient education,” Zaretsky explains.
They also used a question-and-answer format, putting questions in bold type. “We tried to reduce the number of any technical language and acronyms in the vernacular, and we engineered it so that if it’s important for the information to be there, then it is explained,” Zaretsky adds. “If there is a technical word or acronym included in the patient-friendly language, we explain what it means.”
For example, a patient-friendly discharge summary may include this language:
• “Why was I hospitalized? You were hospitalized because of STEMI, which is a severe heart attack. You also had mild protein-calorie malnutrition, which means your body didn’t get enough nutrients from food.”1
The AI-generated summaries were reviewed by physicians, who looked for accuracy and completeness. They also used the Flesch-Kincaid Grade Level and the Patient Education Materials Assessment Tool to assess readability and understandability.
“We hope in the future to test it with actual patients,” Zaretsky says. “We are still in the process of testing and making sure we always have a physician in the loop in the future. We were happy with our results, but as with any artificial intelligence in healthcare, safety is incredibly important. For the foreseeable future, any plans to use it with patients would require physician oversight.”
Zaretsky and colleagues also assessed the AI-generated discharge summaries for a problem that has cropped up in AI writing: hallucinations. “Hallucination” is a term the AI world uses to signify that information is false and not generated or based on any input, Zaretsky explains.
“It can often be stated with seeming confidence,” he adds. “We looked at accuracy in this study, and most of the inaccuracies created were not because of hallucinations, but because of omission. This means the person who evaluated the output felt that there was something in the discharge summary that should have been included in the patient-friendly summary but was not — like a blood test or something.”
This is part of the nuance of using AI-generated writing. It is like having a dial where one could make the writing incredibly comprehensive but have more limited patient-friendly material, Zaretsky says. Or the dial could be turned up on patient-friendly words and be less comprehensive — not using every single detail.
Prospective patients who receive AI-generated discharge summaries will have access to all the materials and information stored in their electronic health records because it is a federal mandate. They will see physician notes and more technical terminology there.
From a case manager’s perspective, AI-generated discharge summaries make their job easier during care transitions. They must assist providers in communicating a lot of information to patients, and these patient-friendly summaries can help them ensure patients understand what their physicians tell them, Zaretsky says.
This is a case where AI can create something useful that does not currently exist for most patients.
“It helps patients understand their care and diagnosis,” Zaretsky says. “In literature, there was one study a decade ago where they called patients after a heart attack, and a huge portion of them didn’t know what their diagnosis was.”
Other methods of explaining discharge summaries to patients are labor intensive, such as employing medical school students on a hotline to answer questions and interpret notes.
“Otherwise, there hasn’t been much. Patients receive discharge instructions and have access to their electronic medical records, but they don’t have a patient-friendly discharge summary document,” Zaretsky says. “That is the potential of AI to not only create this product but to create it at scale.”
Used safely and correctly, AI can create workflow efficiencies and improve healthcare for patients.
“AI is a tool like any other, and it’s something we have to test and something that I think has a lot of potential,” Zaretsky explains. “But we have to do it safely — and that’s the key to all of this.”
In the case of an AI-generated discharge summary, this will not displace anyone’s job because the AI helps to create a new product.
“It’s not taking something else that exists and automating it — it’s creating something that would not exist otherwise,” Zaretsky says. “I always like to be optimistic about these things and think [an AI-generated, patient-friendly discharge summary] would be such a benefit to patients and help them understand their own health. It’s an exciting future and could be a real boon to patient health.”
REFERENCE
- Zaretsky J, Kim JM, Baskharoun S, et al. Generative artificial intelligence to transform inpatient discharge summaries to patient-friendly language and format. JAMA Netw Open 2024;7:e240357.
Researchers found success using artificial intelligence to create a new care transition product — an easy-to-read discharge summary. The new product took the writing to a 6th-grade level from an 11th-grade level and scored higher on a key assessment tool.
You have reached your article limit for the month. Subscribe now to access this article plus other member-only content. TESTING
- Award-winning Medical Content
- Latest Advances & Development in Medicine
- Unbiased Content