Automated detection of gastroesophageal adenocarcinoma recurrence from electronic health records

In gastroesophageal adenocarcinoma (GEA), recurrence following curative-intent resection represents one of the most clinically important outcomes. However, recurrence information is not consistently captured in structured fields within electronic health records (EHRs), and retrospective identification typically relies on labor-intensive manual chart review. We conducted a retrospective, single-center cohort study of patients with GEA who underwent R0 gastrectomy or esophagectomy from July 2020 to October 2024. A fully automated, rule-based natural language processing recurrence detection algorithm was developed using pathology and radiology reports extracted from the EHR. Definitive diagnostic statements in pathology reports were prioritized, while imaging-based recurrence required definitive diagnostic statements at ≥2 time points 14-120 days apart. Algorithm performance was evaluated against manual chart review, which served as the reference standard for recurrence identification. Among 414 eligible patients, 126 (30.4%) experienced recurrence, with 3-year overall survival of 80.3% and recurrence-free survival of 59.3%. The most frequent sites of recurrence were distant lymph nodes, liver, and peritoneum. The algorithm demonstrated high discrimination, with a sensitivity of 90.5%, specificity of 95.1%, positive predictive value of 89.1%, negative predictive value of 95.8%, and accuracy of 93.7%. The median absolute difference in recurrence timing was 0 days, and site classification concordance was 91.2%. This proof-of-concept study demonstrates that postoperative recurrence in GEA can be identified with reasonable reliability using an automated, interpretable algorithm based solely on pathology and radiology reports. Automated recurrence detection may substantially reduce reliance on manual chart review and facilitate scalable real-world effectiveness research and long-term outcome evaluation.