Application of Large Language Models in Stroke Rehabilitation Health Education: 2-Phase Study

Abstract

Background: Stroke is a leading cause of disability and death worldwide, with home-based rehabilitation playing a crucial role in improving patient prognosis and quality of life. Traditional health education often lacks precision, personalization, and accessibility. In contrast, large language models (LLMs) are gaining attention for their potential in medical health education, owing to their advanced natural language processing capabilities. However, the effectiveness of LLMs in home-based stroke rehabilitation remains uncertain.

Objective: This study evaluates the effectiveness of 4 LLMs-ChatGPT-4, MedGo, Qwen, and ERNIE Bot-selected for their diversity in model type, clinical relevance, and accessibility at the time of study design in home-based stroke rehabilitation. The aim is to offer patients with stroke more precise and secure health education pathways while exploring the feasibility of using LLMs to guide health education.

Methods: In the first phase of this study, a literature review and expert interviews identified 15 common questions and 2 clinical cases relevant to patients with stroke in home-based rehabilitation. These were input into 4 LLMs for simulated consultations. Six medical experts (2 clinicians, 2 nursing specialists, and 2 rehabilitation therapists) evaluated the LLM-generated responses using a Likert 5-point scale, assessing accuracy, completeness, readability, safety, and humanity. In the second phase, the top 2 performing models from phase 1 were selected. Thirty patients with stroke undergoing home-based rehabilitation were recruited. Each patient asked both models 3 questions, rated the responses using a satisfaction scale, and assessed readability, text length, and recommended reading age using a Chinese readability analysis tool. Data were analyzed using one-way ANOVA, post hoc Tukey Honestly Significant Difference tests, and paired t tests.

Results: The results revealed significant differences across the 4 models in 5 dimensions: accuracy (P=.002), completeness (P<.001), readability (P=.04), safety (P=.007), and humanity (P<.001). ChatGPT-4 outperformed all models in each dimension, with scores for accuracy (mean 4.28, SD 0.84), completeness (mean 4.35, SD 0.75), readability (mean 4.28, SD 0.85), safety (mean 4.38, SD0.81), and user-friendliness (mean 4.65, SD 0.66). MedGo excelled in accuracy (mean 4.06, SD 0.78) and completeness (mean 4.06, SD 0.74). Qwen and ERNIE Bot scored significantly lower across all 5 dimensions than ChatGPT-4 and MedGo. ChatGPT-4 generated the longest responses (mean 1338.35, SD 236.03) and had the highest readability score (mean 12.88). In the second phase, ChatGPT-4 performed the best overall, while MedGo provided the clearest responses.

Conclusions: LLMs, particularly ChatGPT-4 and MedGo, demonstrated promising performance in home-based stroke rehabilitation education. However, discrepancies between expert and patient evaluations highlight the need for improved alignment with patient comprehension and expectations. Enhancing clinical accuracy, readability, and oversight mechanisms will be essential for future real-world integration.

Keywords: artificial intelligence; health education.; home rehabilitation; large language models; stroke.

Figure 1.. Research design workflow diagram.

Figure 2.. Line chart of the median scores and radar chart for the 4 large language models. (A) Accuracy, (B) completeness, (C) humanity, (D) readability, (E) safety, and (F) radar chart.

Figure 3.. Comparative evaluation of large language model (LLM) responses on relevant questions. (A) Box plot showing the variation in text length among the 4 LLMs, with a significant difference observed between ChatGPT-4 and Qwen (P<.001). (B) Box plot illustrating the variation in reading difficulty scores among the 4 LLMs. (C) Box plot showing the variation in recommended reading age among the 4 LLMs (P=.07). (D) Density plot displaying the distribution of reading difficulty scores among the models. (E) Bar chart showing the distribution of educational levels required to comprehend the responses. P values indicate pairwise comparisons of Chinese character count: ChatGPT-4 versus MedGo (P=.002), ChatGPT-4 versus Qwen (P<.001), ChatGPT-4 versus ERNIE Bot (P=.02), and MedGo versus Qwen (P=.04); and for reading difficulty score: ChatGPT-4 versus MedGo (P=.01). All rating data in this study were tested and found to follow a normal or approximately normal distribution.

Figure 4.. The Sankey diagram illustrates the classification of questions in both phases. On the left, 90 questions posed by 30 patients are shown, while on the right, the 15 integrated questions are displayed.