GPT-4 Improves Readability of Institutional Heart Failure Patient Education Materials
King, R. C.; Samaan, J. S.; Haquang, J.; Bharani, V.; Margolis, S.; Srinivasan, N.; Peng, Y.; Yeo, Y. H.; Ghashghaei, R.
Show abstract
IntroductionHeart failure management involves comprehensive lifestyle modifications such as daily weights, fluid and sodium restriction, and blood pressure monitoring placing additional responsibility on patients and caregivers with successful adherence often requiring extensive counseling and understandable patient education materials (PEMs). Prior research has shown PEMs related to cardiovascular disease often exceed the American Medical Associations 5th-6th grade recommended reading level. The large language model (LLM) Chat Generative Pre-trained Transformer (ChatGPT) may be a useful tool for improving PEM readability. Materials and MethodsA total of 143 heart failure PEMs were collected from the websites of the top 10 institutions listed on the 2022-2023 US News & World Report for "Best Hospitals for Cardiology, Heart & Vascular Surgery". PEMs were individually entered into GPT-4 (Version updated 20 July 2023) preceded by the prompt "please explain the following in simpler terms". The readability of the institutional PEM and ChatGPT revised PEM were both assessed using Textstat library in Python and the Textstat readability package in R software. The accuracy and comprehensiveness of revised GPT-4 PEMs were assessed by a board-certified cardiologist. ResultsThe Flesch-Kincaid grade reading level ranged from 8th grade to college freshman with a median of 10th grade vs 6th to 8th grade with a median of 7th grade for institutional PEMs and GPT-4 PEMs (p< 0.001), respectively. There were 13/143 (9.1%) institutional PEMs below the 6th grade reading level which improved to 33/143 (23.1%) after revision by GPT-4 (p<0.001). No GPT-4 revised PEMs were graded as less accurate or less comprehensive compared to institutional PEMs. A total of 33/143 (23.1%) GPT-4 PEMs were graded as more comprehensive. ConclusionsGPT-4 significantly improved the readability of institutional heart failure PEMs. The model may be a promising adjunct resource in addition to care provided by a licensed healthcare professional for patients living with heart failure. Further rigorous testing and validation is needed to investigate its safety, efficacy and impact on patient health literacy.
Matching journals
The top 7 journals account for 50% of the predicted probability mass.
Similar papers in this journal
- Cardiology Knowledge Assessment of Retrieval-Augmented Open versus Proprietary Large Language Models 94%
- Theory of radiologist interaction with instant messaging decision support tools: a sequential-explanatory study 93%
- Ethical review of clinical research with generative AI: Evaluating ChatGPT’s accuracy and reproducibility 93%
Similar papers in this journal
- Integrating remote monitoring into heart failure patients’ care regimen: A pilot study 94%
- Clinical code sets and the problem of redundancy in code set repositories 93%
- Evaluating user experience with immersive technology in simulation-based education: a modified Delphi study with qualitative analysis 93%
Similar papers in this journal
Similar papers in this journal
- A digital self-care intervention for Ugandan patients with heart failure and their clinicians: User-centred design and usability study 93%
- Validating a Clinical Decision Support System for Palliative Care using healthcare professionals’ insights 93%
- How suitable are clinical vignettes for the evaluation of symptom checker apps? A test theoretical perspective 92%
Similar papers in this journal
- Machine learning approaches to predict 30-day mortality following percutaneous coronary intervention in an Australian population 92%
- The Impact of COVID-19 Pandemic on Cardiology Services 91%
- Multispecialty multidisciplinary input into comorbidities in heart failure reduces hospitalisation and clinic attendance 90%
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.