Prognostic model using LODDS for EGJ carcinomas based on the SEER database: the development of an interactive online tool

Background and aims: The American Joint Committee on Cancer staging is used worldwide to predict cancer prognoses, but we found that it cannot clearly stratify stage III and stage IV esophagogastric junction (EGJ) carcinoma, which is difficult to predict due to its distinct anatomical location and behavior. We developed and tested a prognostic model for patients with EGJ carcinoma using the Surveillance, Epidemiology, and End Results (SEER) database, and we constructed an online visualization tool for individualized risk assessment. Materials and methods: We collected the clinical data of patients with malignant tumors of the EGJ from the SEER database and an independent single-center cohort to form the external validation cohort. Variables were screened using the Boruta algorithm, Spearman’s method, and the variance inflation factor test. Seven survival models were established: Cox proportional hazards (CoxPH), ridge regression, elastic net, random survival forest, extreme gradient boosting, gradient boosting machine (GBM), and conditional inference tree. The models’ performances were compared using the C-index, the time-dependent area under the receiver operating characteristic curve, calibration curves, and a decision curve analysis. Results: The results identified patient age, tumor differentiation, T-stage, tumor size, log odds of lymph node positivity (LODDS), chemotherapy, and histological subtype as independent prognostic factors. Among the constructed models based on these variables, the CoxPH model and the GBM model demonstrated good robustness and generalizability. Their C-index values in the training, internal validation, and external validation cohorts were 0.703, 0.676, and 0.739, and 0.712, 0.678, and 0.736, respectively. Considering its high clinical interpretability, we developed an online interactive web calculator based on the CoxPH model using the R Shiny framework to enable individualized survival prediction. Conclusion: The CoxPH model and the GBM model, which incorporate core pathological indicators such as LODDS, demonstrated consistent predictive performance in forecasting survival outcomes for patients with malignant EGJ tumors. The online tool provides clinicians with personalized predictions of 12-, 36-, and 60-month survival probabilities and holds potential clinical value; however, its clinical utility requires further validation in additional prospective multicenter cohorts.