Evaluating reliability of large language models for patient queries: Concordance with NCCN invasive breast cancer guideline using ChatGPT, DeepSeek, and Gemini.

Abstract

e12550 Background: The increasing use of large language models (LLMs) by patients seeking healthcare information emphasizes the need to evaluate their clinical reliability. Invasive breast cancer, characterized by nuanced, multi-stage treatment protocols, represents a key domain where LLM-generated patient guidance must be assessed against established clinical standards. Methods: We developed 62 patient-oriented questions based on NCCN guidelines for patients on invasive breast cancer, spanning eight clinical domains: general information, testing, staging, types of treatment, postoperative outcomes, supportive care, treatment options, and recurrence. Responses were generated by ChatGPT (GPT-5.2), DeepSeek (V3.2), and Gemini (2.5 Flash). Three board-certified oncologists evaluated each response across five metrics—factual accuracy, clarity, coherence, relevance, and completeness—using a 5-point Likert scale. Inter-rater reliability was assessed with a two-way random-effects intraclass correlation coefficient (ICC2). Differences between models were analyzed using Friedman’s chi-square test and Wilcoxon signed-rank post-hoc tests, with descriptive statistics computed per domain. Results: Inter-rater reliability was moderate for completeness (ICC = 0.39) and low for other metrics (ICC = 0.09–0.26). Across 186 responses (62 questions × 3 models), DeepSeek led in completeness (mean 4.65) and factual accuracy (4.26), Gemini in coherence (4.79) and relevance (4.56), and ChatGPT in clarity (4.84). Friedman tests showed significant differences in completeness (χ² = 42.05, p < 0.001) and clarity (χ² = 19.12, p < 0.001). Holm-adjusted Wilcoxon tests confirmed DeepSeek > ChatGPT in completeness (p < 0.001), Gemini > DeepSeek in coherence (p = 0.048), and ChatGPT > DeepSeek in clarity (p < 0.001). Clinical domain analyses revealed variability: DeepSeek excelled in completeness for supportive care (4.73/5, n = 15) and treatment options (4.50/5, n = 10), outperforming ChatGPT (3.53 & 3.60) and Gemini (4.60 & 4.10). Conversely, ChatGPT and Gemini were superior in clarity/coherence for informational domains—general information about invasive breast cancer (both 5.00 clarity, 5.00/4.86 coherence, n = 7 vs DeepSeek 3.71) and testing for breast cancer (both ~5.00/4.80–5.00, n = 5 vs DeepSeek 4.60). This pattern shows the need for domain-tuned LLMs in clinical patient guidance. Conclusions: The quality of LLM-generated patient information is highly domain-dependent, with each model exhibiting distinct strengths. Low inter-rater reliability emphasizes the challenge of consistent evaluation. While LLMs can serve as supplementary educational tools, their clinical integration requires careful, domain-specific validation and clear guidance regarding their variable performance across different topics in breast cancer care.