Abstract
BACKGROUND: Professional opinion polling has become a popular means of seeking advice for complex nephrology questions in the #AskRenal community on X. ChatGPT is a large language model with remarkable problem-solving capabilities, but its ability to provide solutions for real-world clinical scenarios remains unproven. This study seeks to evaluate how closely ChatGPT's responses align with current prevailing medical opinions in nephrology. METHODS: Nephrology polls from X were submitted to ChatGPT-4, which generated answers without prior knowledge of the poll outcomes. Its responses were compared to the poll results (inter-rater) and a second set of responses given after a one-week interval (intra-rater) using Cohen's kappa statistic (κ). Subgroup analysis was performed based on question subject matter. RESULTS: Our analysis comprised two rounds of testing ChatGPT on 271 nephrology-related questions. In the first round, ChatGPT's responses agreed with poll results for 163 of the 271 questions (60.2%; κ = 0.42, 95% CI: 0.38-0.46). In the second round, conducted to assess reproducibility, agreement improved slightly to 171 out of 271 questions (63.1%; κ = 0.46, 95% CI: 0.42-0.50). Comparison of ChatGPT's responses between the two rounds demonstrated high internal consistency, with agreement in 245 out of 271 responses (90.4%; κ = 0.86, 95% CI: 0.82-0.90). Subgroup analysis revealed stronger performance in the combined areas of homeostasis, nephrolithiasis, and pharmacology (κ = 0.53, 95% CI: 0.47-0.59 in both rounds), compared to other nephrology subfields. CONCLUSION: ChatGPT-4 demonstrates modest capability in replicating prevailing professional opinion in nephrology polls overall, with varying performance levels between question topics and excellent internal consistency. This study provides insights into the potential and limitations of using ChatGPT in medical decision making.