To identify distinct clinical and endoscopic phenotypes among patients presenting with non-variceal upper gastrointestinal bleeding (NVUGIB), based on age, sex, clinical presentation, and exposure to antithrombotic or NSAID therapy, using an unsupervised machine-learning clustering technique.

We performed an unsupervised clustering analysis using Partitioning Around Medoids (PAM) on a Gower distance matrix. Two-dimensional visualization of the cluster structure was obtained with t-distributed Stochastic Neighbor Embedding (t-SNE). The analysis included 405 patients ≥16 years old hospitalized with clinical signs of NVUGIB, without history of gastrointestinal malignancy, portal hypertension, or varices, at the General Hospital of Ioannina, Greece.

PAM clustering (k = 3) identified three distinct clinical and endoscopic phenotypes among these patients. Cluster 1 included only elderly female patients, most of whom were aged 80 years or older. Nearly all of these patients presented with melena. Endoscopic evaluations revealed no identifiable source of bleeding in 73% of the cases. The remaining patients showed a mix of low-frequency lesions, such as arteriovenous malformations, Dieulafoy lesions, duodenal ulcers, esophagitis, Mallory-Weiss tears, and gastric ulcers. Despite the lack of clear focal endoscopic findings that could explain the suspected bleeding, patients in this cluster had significant exposure to direct oral anticoagulants (DOACs), along with smaller contributions from heparin, salospir, and dual antiplatelet therapy. This pattern indicates subtle mucosal bleeding caused by anticoagulants, which may result in melena without distinct endoscopic lesions. Cluster 2 included predominantly male patients across a wide age range, including the youngest individuals in the cohort. Drug exposure was limited, with most patients not receiving antiplatelet or anticoagulant therapy, although NSAID and salospir use were present. This group demonstrated the highest prevalence of gastric and duodenal ulcers, forming an “ulcer-driven” bleeding phenotype, likely reflecting underlying peptic ulcer disease rather than drug-related injury. Cluster 3 consisted of very elderly male patients, mostly aged 80 and older, who had significant exposure to DOACs and smaller contributions from heparin, salospir, and Dual Antiplatelet Therapy. Melena was the dominant mode of presentation. Most endoscopies demonstrated no relevant bleeding-related findings, while the remainder revealed only low-frequency lesions, primarily duodenal ulcers, arteriovenous malformations, and esophagitis. This phenotype reflects bleeding presentations in elderly men receiving DOAC therapy, in whom mucosal bleeding is often subtle, diffuse, or endoscopically inapparent despite clinically evident melena.

This analysis identified three distinct phenotypes of NVUGIB based on age, sex, and anticoagulant exposure. Elderly patients on anticoagulants often presented with melena but lacked identifiable endoscopic lesions, suggesting subtle mucosal bleeding. Conversely, patients with minimal drug exposure showed ulcer-related bleeding. These findings underscore the diversity of NVUGIB presentations. Clustering patient profiles may help endoscopists to anticipate endoscopic yield and prepare more effectively for managing potential findings, thereby optimising the endoscopic management of lesions.