Feeding intelligence: comparative evaluation of ChatGPT and clinical guidelines for nutritional management in head and neck cancer

Abstract

Artificial intelligence (AI) tools such as ChatGPT are increasingly applied in digital health and patient education, yet their alignment with established clinical guidelines for cancer-related nutritional management remains unclear. This study aimed to evaluate the concordance, functional characteristics, patient accessibility, and innovation of ChatGPT-generated nutritional recommendations compared with clinical guidelines from the Chinese Society of Clinical Oncology (CSCO), Chinese Nutrition Society (CNS), and European Society for Clinical Nutrition and Metabolism (ESPEN). We analyzed ChatGPT responses across six key nutrition-related issues—anorexia/cachexia, dysphagia, oral mucositis, unintentional weight loss, gastrointestinal intolerance, and nutritional monitoring—and compared them with guideline recommendations. Expert evaluation (n = 5), readability metrics, semantic similarity (TF-IDF), and patient-centered assessments were conducted to compare personalization, innovation, clinical feasibility, evidence-based support, population applicability, clarity, and self-management guidance. ChatGPT recommendations aligned with at least one guideline in 50.0–64.3% of cases, highest for dysphagia (64.3%), and included general strategies such as small frequent meals, texture modification, hydration, and high-protein/high-calorie intake. ChatGPT-specific suggestions (8.3–18.2%) focused on lifestyle and behavioral interventions, including mindful eating, music therapy, and wearable diet trackers. Expert ratings indicated higher personalization (4.3/5) and innovation (4.6/5) for ChatGPT, whereas guidelines scored higher for clinical feasibility (4.7/5), evidence-based support (4.9/5), and population applicability (4.8/5). Between-group differences were statistically significant for clinical feasibility, evidence-based support, and applicability (all p < 0.01; 95% CI for mean differences: 0.62–1.12), whereas personalization showed no significant difference (p = 0.063). ChatGPT exhibited superior patient-centered performance in clarity (4.5 vs. 3.2, p = 0.004, 95% CI: 0.47–2.13) and self-management guidance (4.6 vs. 3.0, p = 0.002, 95% CI: 0.65–2.05) and demonstrated more concise, readable content (Flesch–Kincaid grade 12.9–14.2) compared with guidelines (17.9–20.5). Semantic analysis revealed moderate overlap with CSCO (≈ 0.63) and CNS (≈ 0.59), and lower similarity with ESPEN (≈ 0.47), highlighting ChatGPT’s use of patient-friendly language. Topic modeling identified three clusters: patient support and accessibility (ChatGPT), technical nutrition therapy (ESPEN/CSCO), and nutritional assessment and monitoring (CNS). ChatGPT provides personalized, innovative, and patient-accessible nutritional guidance for cancer-related malnutrition, complementing traditional clinical guidelines. While guidelines remain essential for evidence-based decision-making, AI tools may enhance patient education, engagement, and self-management in digital health applications.

Ähnliche Arbeiten