GPT-5.2 vs. DeepSeek-V3.2 in Simplifying Cardiac Magnetic Resonance Reports: A Prospective Real-World Study

Abstract

The aim of this study was to assess the feasibility of two large language models (LLMs), GPT-5.2 and DeepSeek-V3.2, for simplifying cardiac magnetic resonance (CMR) reports into participant-accessible language. Participants undergoing CMR examinations were prospectively recruited. Original reports were randomly assigned in a 1:1 ratio to either GPT-5.2 or DeepSeek-V3.2. Predesigned prompts were used to guide the LLMs in generating simplified reports. Two customized structured Likert-scale questionnaires were developed to assess the performance and comprehensibility of the LLM-generated reports. The internal consistency and factor structure of these questionnaires were evaluated. A total of 117 participants were initially recruited. After excluding four cases with unacceptable LLM-generated reports (3 numerical errors and 1 severity misclassification), seven who declined to complete the questionnaire, and six with incomplete responses, 100 participants were included in the final analysis (mean age 48.8 ± 12.6 years; 72% male). Both the performance and comprehension questionnaires demonstrated good internal consistency and interpretable factor structures. The simplified reports were associated with significantly higher questionnaire-assessed participant comprehension scores than the original reports across all four dimensions (all P < 0.013). No statistically significant differences were observed between GPT-5.2 and DeepSeek-V3.2 in radiologist-rated performance (all P > 0.01) or questionnaire-assessed participant comprehension (all P > 0.013), and no significant agreement or association was observed between LLM-rated and radiologist-rated scores (all P > 0.05). LLMs are a promising tool for translating CMR reports into participant-accessible language. GPT-5.2 and DeepSeek-V3.2 showed no significant difference in questionnaire-assessed participant comprehension or radiologist-rated performance. However, radiologist supervision remains necessary to ensure the quality and reliability of LLM-generated reports.