A new editorial in JGH Open takes a close look at GastroGPT, a specialty-trained LLM built specifically for gastroenterology. Trained on 1.2M GI-specific tokens — including guidelines, peer-reviewed literature, and 10,000 synthetic vignettes — the model outperformed general LLMs (GPT-4, Bard, Claude) across six of seven clinical tasks, from history-taking to referral decisions to patient education.
Early results are impressive: in simulated IBD, endoscopy, and hepatology cases, GastroGPT scored 8.1/10, with strong reproducibility across case complexity. But the authors also flag real-world limitations — simulated datasets, specialty-biased training, and the need for comparisons with purpose-built medical AI models.