Patient Comprehension and Perception of AI-Generated Clinical Notes in Musculoskeletal Healthcare: An Experimental Proof-of-Concept Study
Translated title
Patienters forståelse og opfattelse af AI-genererede journalnotater inden for det muskuloskeletale sundhedsområde: Et eksperimentelt proof-of-concept-studie
Authors
Keskin, Daniel ; Hansen, Mathias Louie ; Sand, Katrine Vilja
Term
4. term
Education
Publication year
2025
Submitted on
2025-06-02
Pages
71
Abstract
Omkring 1,71 milliarder mennesker globalt lever med muskel- og skeletlidelser (problemer i muskler, knogler og led), som giver funktionsnedsættelse, lavere livskvalitet og høje sundhedsudgifter. Det er vigtigt at kunne forstå, hvad klinikere skriver i elektroniske patientjournaler (EHR), men kliniske notater er ofte fulde af fagsprog. Det er særligt udfordrende for personer med begrænset sundhedskompetence (evnen til at finde, forstå og bruge sundhedsinformation). Nye generative AI-systemer kan omskrive notater til mere enkel, patientvenlig tekst. Dette proof-of-concept-studie undersøgte, om AI-genererede, patientvenlige versioner af kliniske notater om muskel- og skeletlidelser forbedrer forståelsen og deltagernes oplevelse sammenlignet med de originale notater, og om effekten varierer med sundhedskompetence. 19 deltagere (gennemsnitsalder 55,7 år, SD 19,9) læste seks kliniske notater i både original og AI-forenklet version. De forenklede notater blev skabt med GPT-4o ved hjælp af zero-shot prompting (instruktioner uden eksempler) tilpasset personer med lav sundhedskompetence. Objektiv forståelse blev målt med notespecifikke tjeklister, og oplevelse blev målt med et spørgeskema med seks spørgsmål på en fem-punkt Likert-skala. Sundhedskompetence blev målt med de danske HLS-EU-Q16 og DS-TOFHLA. Statistiske analyser omfattede ikke-parametriske tests, parrede t-tests og lineære mixed models. Deltagerne opnåede signifikant højere objektiv forståelsesscorer for AI-genererede notater (median 80 %, IQR 44 %) end for originale notater (median 38 %, IQR 44,5 %; z = -3,823, P < .001). Subjektive vurderinger favoriserede de AI-genererede notater på alle dimensioner (alle P < .001). Der var en signifikant interaktion mellem sundhedskompetence og noteversion, med større forståelsesgevinster for personer med lavere sundhedskompetence (F(1, 85,8) = 8,9, P = .004). Resultaterne peger på, at AI kan mindske forståelseskløften og styrke patienternes handlekraft. En integration af AI-baseret forenkling af notater i rutinepraksis kan øge patientinddragelse. Fremtidige studier bør udvikle og validere værktøjer, der specifikt måler patienters forståelse af kliniske notater.
About 1.71 billion people worldwide live with musculoskeletal disorders (problems of muscles, bones, and joints), leading to disability, lower quality of life, and high healthcare costs. Being able to clearly understand what clinicians write in electronic health records (EHRs) is important, but clinical notes are often full of medical jargon. This is especially challenging for people with limited health literacy (the ability to find, understand, and use health information). New generative AI systems can rewrite notes in simpler, more patient-friendly language. This proof-of-concept study tested whether AI-generated, patient-friendly versions of musculoskeletal clinical notes improve understanding and how people feel about the notes, compared with the original versions, and whether effects differ by health literacy. Nineteen participants (mean age 55.7 years, SD 19.9) read six clinical notes in their original form and in AI-generated simplified form. The simplified notes were produced with GPT-4o using zero-shot prompting tailored to people with low health literacy. Objective comprehension was measured with checklists created for each note, and perceptions were measured with a six-item questionnaire on a five-point Likert scale. Health literacy was measured with the Danish HLS-EU-Q16 and DS-TOFHLA instruments. Statistical analyses included nonparametric tests, paired t-tests, and linear mixed models. Participants scored significantly higher on objective comprehension for AI-generated notes (median 80%, IQR 44%) than for original notes (median 38%, IQR 44.5%; z = -3.823, P < .001). Subjective ratings favored the AI-generated notes across all dimensions (all P < .001). There was a significant interaction between health literacy and note version, with greater comprehension gains for individuals with lower health literacy (F(1, 85.8) = 8.9, P = .004). These findings support AI as a tool to bridge understanding gaps and empower patients. Integrating AI-driven note simplification into routine clinical practice could enhance patient engagement. Future studies should prioritize developing and validating tools specifically designed to measure patients’ comprehension of clinical notes.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents