Abstract
BACKGROUND: Patients often struggle to understand standard hospital discharge letters, increasing the risk of medication errors and misunderstandings. According to cognitive load theory (CLT), complex, information-dense texts can overload working memory and impair comprehension. Artificial intelligence tools that generate patient-centered versions could help reduce extraneous cognitive load and bridge this gap. However, evidence for their effectiveness remains limited. OBJECTIVE: This study aimed to evaluate whether GPT-4 (OpenAI)-generated patient-centered letters improve standardized patients' retention and understanding of safety-relevant medical information compared with standard hospital discharge letters, and to explore potential effects on cognitive load as described by CLT. METHODS: In this prospective, randomized, crossover study, 48 trained standardized patients received a conventional discharge letter for an assigned disease (out of 3) and its matching GPT-4-generated patient-centered letter. Participants read one version first, identified predefined safety-relevant "learning objectives," and then repeated the task with the alternate version. The primary outcome was the proportion of learning objectives fully, partially, or not reported. In a secondary analysis, results were stratified by content field (Medication, Organization, Prevention of Complications, Lifestyle/Disease Management) and Bloom taxonomy level ("Remember," "Understand"). RESULTS: The letter type significantly influenced comprehension (odds ratio [OR] 1.74, 95% CI 1.45-2.08; P<.001). Patient letters, compared with discharge letters, led to higher rates of fully (490/1073, 45.7% vs 413/1073, 38.5%) or partially (322/1073, 30% vs 287/1073, 26.7%) stated learning objectives and fewer omissions (261/1073, 24.3% vs 373/1073, 34.8%). Participants performed better on "Remember" than on "Understand" learning objectives, regardless of letter type (OR 3.33, 95% CI 1.96-5.88; P<.001). Compared with standard hospital discharge letters, patient letters consistently improved results at both cognitive levels ("Remember": 278/545, 51% vs 242/545, 44.4%; "Understand": 212/528, 40.2% vs 171/528, 32.4% fully stated). The effect of patient letters varied by content field (P<.001). The greatest improvements were observed for "Medication" (170/254, 66.9% vs 129/254, 50.8% fully stated) and "Organization" (78/158, 49.4% vs 62/158, 39.2% fully stated). Improvements in the content field "Prevention of Complications" were modest, and those for "Lifestyle/Disease Management" were even smaller across all conditions. A total of 24.3% (261/1073) of key information remained unrecognized. CONCLUSIONS: In this explanatory study, GPT-4-generated patient letters improved comprehension of safety-relevant discharge information among standardized patients, particularly regarding medication and organizational aspects. However, they were less effective in supporting higher-order understanding, such as risk prevention or lifestyle management. These hypothesis-driven findings can be interpreted within a CLT framework and may motivate prospective evaluation of multimodal, iterative supports.