Speakers

Anivarya Kumar, BA
Fourth-Year Medical Student
Duke University School of Medicine

Matthew Engelhard, MD, PhD
Assistant Professor, Department of Biostatistics & Bioinformatics
Duke University School of Medicine

Slides

Keywords

Health Literacy; Large Language Models; Artificial Intelligence; Electronic Health Records

Key Points

• Limited health literacy (HL) has tangible effects on morbidity and mortality: it’s associated with higher rates of hospital admissions and readmissions; medication nonadherence; healthcare costs; and all-cause mortality. 9 in 10 adults have limited HL, and rates are 2 – 3 times lower in marginalized populations.

• 71% of patients report accessing their electronic health records (EHRs) to read documentation from their clinical visits, particularly the discharge summary notes (DSNs). But clinical notes have low levels of readability, hindering patients’ ability to engage in shared decision-making.

• The research team looked at whether a Generative-Pre-trained-Transformer-4 (GPT-4)-based plain language translation of DSNs could improve patient comprehension of disease course and management.

• 533 patients, recruited from a pool of EHR users, were randomly assigned 4 DSNs to assess. After reading the DSNs – 2 translated into more accessible language, 2 untranslated – patients answered questions assessing their objective comprehension, subjective comprehension, confidence, and time spent on each DSN.

• Compared to the untranslated DSNs, objective understanding of the translated DSNs increased by 6.1%; subjective understanding increased 18%; confidence increased 45%; and average time spent with the DSNs decreased 51%.

• The research team concluded that GPT translation of DSNs significantly improved patient comprehension of disease course and management and optimized time spent reading them. The effect was significantly greater in marginalized populations with historically low health literacy, reducing the gap in comprehension scores between patient populations.

• Limitations included the use of standardized DSNs as opposed to real-world DSNs; the use of MyChart when enrolling patients, leading to a participant group with a higher baseline HL; and the modest number of Hispanic patients enrolled in the study.

• Race is a significant and independent factor for HL. Preliminary data suggests that GPT translation can help close this gap. The research team identified this as an area for further study.

Discussion Themes

While discharge instructions alone can be great for providing patients with action items, they lack some of the context that DSNs can provide, lending the patient a more complete understanding of their condition.

The advantages of providing pre-generated materials, as opposed to pointing patients to an large language model (LLM) like Chat GPT for a more interactive explanation of their condition, include the potential for screening by a healthcare professional and less of a burden on the patient.

The study team ended up favoring “semantically-focused” translations over translations that focused solely on simplifying the language or avoiding jargon. When the LLM was asked to focus on semantics, it was more likely to define concepts and their implications.

Health literacy and reading level are not necessarily on par, and patient-centric or accessible language/LLMs are very important to consider. This may require further investigation, e.g. through qualitative interviews.