Management of Upper Gastrointestinal Bleeding: AI, A Decision-Making Tool?

Aims The objective of this study is to assess the effectiveness of AI in triaging upper gastrointestinal bleeding cases in emergency settings, by identifying patients who require immediate or delayed esophagogastroduodenoscopy (EGD).

Methods The medical records of 100 patients who presented to the emergency department for UGIB between April 2024 and September 2024 were retrospectively reviewed. Each case was submitted to two AI models: Google GEMINI (September 2024 version) and ChatGPT (version 3.5). As a virtual gastroenterology experts, they were asked whether they would recommend urgent or delayed endoscopy for each case. The clinical data provided for each patient included age, medical history, clinical examination findings, and standard laboratory results at admission. The AI recommendations were recorded in an Excel file and compared to the Glasgow-Blatchford Score (GBS), calculated by the gastroenterologist upon patient admission. An urgent endoscopy was considered necessary for any GBS≥3. Statistical analysis was performed using Jamovi software, version 2.3. The level of agreement between the AI models and the human evaluator was measured using Cohen’s kappa coefficient (κ).

Results Regarding the patients' epidemiological data, the median age was 63 years [37.5; 78], with a male-to-female ratio of 1.7. Of the cohort, 73% had one or more comorbidities: 33% were hypertensive, 22% diabetic, 17% had ischemic heart disease, and 12% had chronic liver disease. At admission, melena was the primary symptom in 78% of patients. The mean GBS was 7.07±3.47. ChatGPT recommended urgent endoscopy in 48% of cases, while Google GEMINI suggested it in 49%. A moderate agreement was observed between the two AI models for urgent endoscopy recommendations (kappa=0.540, p<0.001). However, there was poor agreement between human evaluations and AI recommendations. The agreement between the gastroenterologist and ChatGPT for urgent endoscopy was weak (kappa=0.222, p<0.001), with similar results observed with Google GEMINI (kappa=0.231, p<0.001).

Conclusions Artificial intelligence is a promising tool that is already beginning to revolutionize several sectors of medicine. However, our study shows that currently available public AI models are not yet capable of effectively triaging patients in emergency settings, particularly in cases of upper gastrointestinal bleeding. A model specifically trained on clinical gastroenterology cases, based on validated scientific guidelines, could yield better results and potentially transform patient triage in emergency departments.

Publikationsverlauf

Artikel online veröffentlicht:
27. März 2025