Assessment of accuracy, quality, and readability of large language models in kidney transplantation practice

This study aimed to comparatively evaluate the accuracy, overall quality, and readability of responses generated by three contemporary large language models (LLMs)-Grok AI, ChatGPT-5 Plus, and Gemini 2.5 Flash-to clinical questions commonly encountered in kidney transplantation (KT) practice.

Ten open-ended clinical scenarios covering surgical techniques, complication management, immunosuppressive strategies, and perioperative care were independently assessed by two experienced physicians. LLM-generated responses were evaluated using a structured rubric assessing scientific accuracy, comprehensiveness, clarity, and clinical relevance (equally weighted), along with the Global Quality Score (GQS). Readability was analyzed using the Flesch Reading Ease Score (FRES). Inter-rater reliability was determined using Cronbach's alpha and intraclass correlation coefficients (ICC).

All three LLMs demonstrated consistently high performance, with mean rubric scores exceeding 9.2/10 and GQS values above 4.4/5. Grok AI achieved the highest numerical scores (rubric: 9.5 ± 0.71; GQS: 4.7 ± 0.48), followed by ChatGPT-5 Plus (9.4 ± 0.70; 4.6 ± 0.52) and Gemini 2.5 Flash (9.2 ± 1.03; 4.4 ± 0.70); however, no statistically significant differences were observed (p > 0.05). FRES values ranged from 23.85 to 32.19, indicating professional-level readability suitable for physicians. Inter-rater reliability was high across all measures (Cronbach's alpha: 0.737-0.889; ICC: p < 0.05). Key limitations include the potential for LLM hallucinations, the limited number of scenarios, and the evaluation of text-based responses at a single time point.

Current-generation LLMs show high and comparable performance in addressing complex KT-related clinical questions. Their consistent accuracy, quality, and readability support their potential role as adjunctive clinical decision-support tools, provided appropriate physician oversight.