14 | Application of an artificial intelligence algorithm to patients presenting to the emergency department of the Pisa University Hospital for chest pain: retrospective assessment of diagnostic accuracy

Background. Chest pain is a frequent symptom encountered in the Emergency Department (ED), with a mortality rate of 2-4%. It may reflect a wide range of clinical conditions, from benign issues to life-threatening Cardiovascular (CV) diseases. Alongside conventional clinical tools and various validated scores, Artificial Intelligence (AI) systems are emerging as potential aids for appropriate risk stratification.

Methods. A sample of 154 patients who presented to the ED of the Pisa University Hospital (AOUP) between May 2024 and June 2025 for chest pain was analyzed to evaluate the diagnostic accuracy of the ChatGPT 4.0 algorithm. For each patient, management, final diagnosis, and outcomes were reviewed. A limited number of variables were selected and submitted to the algorithm in two sequential queries: the first included triage data (demographics, sex, vital signs, medical history, and pain characteristics), while the second included physical examination findings, laboratory tests, and imaging results.

Results. The study population had a mean age of 60.7 years. Life-threatening conditions accounted for a clear minority of cases (17.45%). Most patients (74.5%) were discharged home after negative diagnostic evaluations. The results concerning diagnostic accuracy between the final clinical diagnosis and the AI-suggested diagnosis showed, in the first query (triage data), a concordance of 32.9% when considering only the first proposed diagnosis and 67.8% when considering the top three suggestions. In the second query (which incorporated laboratory and instrumental data), accuracy increased to 90.6%. Similarly, the algorithm’s ability to correctly identify exclusively acute CV conditions (pulmonary embolism, acute coronary syndrome, and aortic dissection) reached 100%.

Conclusions. AI demonstrated an ability to recognize relevant diagnostic patterns, particularly when provided with well-structured inputs. The promising results obtained open concrete possibilities for using AI systems in triage and early evaluation of chest pain, potentially leading to purpose-built algorithms capable of integrating directly with hospital information systems.