Abstract DP013: Improving Comprehension of Stroke Health Materials in Cabo Verde Using LLM-Assisted Translation and Simplification

Introduction: Health materials are often written above recommended readability levels, limiting comprehension in multilingual, resource-constrained settings. We evaluated whether large language model (LLM)–assisted translation/simplification improves understanding of short excerpts from World Health Organization (WHO) patient fact sheets on stroke, hypertension, and diabetes among adults in Cabo Verde. Methods: We conducted a multi-site randomized study in primary-care clinics in Cabo Verde; 63 adults were assigned to read WHO fact-sheet excerpts presented as (1) Portuguese Google translated (baseline), (2) Portuguese GPT simplified, or (3) Kriolu GPT simplified. Comprehension was assessed through an excerpt-specific quiz; the primary endpoint was percent correct. Pooled analyses used a two-way ANOVA (language arm; disease topic), with a prespecified stroke-only one-way ANOVA. An education-adjusted ANCOVA served as a sensitivity analysis. Participants also rated ease of understanding and translation accuracy (Likert 1–5), analyzed separately. Results: Across all topics, quiz scores were higher in both LLM arms (baseline 54.3%, n=22; Portuguese GPT 67.4%, n=24; Kriolu GPT 66.3%, n=17). The language main effect was significant (p=0.02), while disease was not (p=0.96). Versus baseline, mean differences were +13.1 points for Portuguese GPT (95% CI +3.5 to +22.8) and +12.1 points for Kriolu GPT (95% CI +1.4 to +22.7). In the stroke-only subset, the pattern persisted (baseline 54.2%, n=11; Portuguese GPT 67.6%, n=8; Kriolu GPT 69.5%, n=4), with pairwise differences of +13.4 (95% CI +1.8 to +25.0) and +15.3 (95% CI +1.8 to +28.8), respectively. The education-adjusted ANCOVA yielded similar conclusions. Subjective ratings were uniformly favorable and did not differ by arm (medians: ease 3/5, accuracy 4/5; both p>0.40). Conclusions: LLM-assisted Portuguese and Kriolu materials significantly improved comprehension, with consistent gains across stroke, hypertension, and diabetes—independent of education status. This language-inclusive approach offers a scalable path to strengthen stroke and non-communicable disease education in low-resource, multilingual settings while advancing patient engagement at the point of care.