Symptom profiles and cancer stage at diagnosis in gastroesophageal junction adenocarcinoma: An ordinal and clustering analysis.

Abstract

e16005 Background: Gastroesophageal junction adenocarcinoma (GEJA) is often diagnosed at advanced stages, contributing to poor outcomes. While individual presenting symptoms have been linked to disease severity, less is known about how combinations of symptoms or symptom profiles relate to stage at diagnosis. Methods: We analyzed 250 patients with GEJA staged I–IV at diagnosis. Analyses were restricted to presenting clinical symptoms (e.g., dysphagia, abdominal pain, systemic weight loss), excluding risk factors and surveillance-related variables. Ordinal logistic regression was used to model associations between symptom presence and increasing stage. Penalized ordinal regression (LASSO) was applied for symptom selection, followed by refitting of a reduced ordinal model. Proportional odds assumptions were assessed using the Brant test. Predicted stage probabilities were generated to visualize symptom specific shifts in stage distribution. In parallel, unsupervised clustering approaches, including symptom prevalence heatmaps, stage-specific enrichment analyses, and patient-level clustering using Jaccard distances were used to identify symptom groupings associated with disease stage. Sensitivity analyses compared early (stage I–III) versus advanced disease (stage IV). Results: LASSO identified dysphagia, abdominal pain, and systemic weight loss as the most informative symptoms. In the refitted ordinal model, all three were independently associated with higher stage at diagnosis (ORs ≈ 2.3–3.1). Predicted probability plots demonstrated substantial increases in the probability of stage IV disease when these symptoms were present. Although proportional odds assumptions were partially violated for select symptoms, effect directions were consistent across sensitivity analyses. Symptom clustering revealed distinct stage-associated profiles, with advanced-stage disease characterized by co-occurrence of obstructive and systemic symptoms, while earlier stages showed lower symptom burden. Conclusions: Symptom-only modeling and clustering helped identify clinically meaningful symptom profiles associated with GEJA stage at diagnosis. Combining ordinal modeling with visualization and unsupervised clustering provides intuitive insight beyond single-symptom analyses and may support earlier recognition of advanced disease in symptomatic patients.