Prediction-model development and temporal validation for six-month cancer-specific mortality in de novo metastatic gastroesophageal junction adenocarcinoma: a SEER-based study.

BACKGROUND: De novo metastatic gastroesophageal junction adenocarcinoma (GEJA) is associated with heterogeneous short-term outcomes. A six-month cancer-specific mortality (CSM) horizon is clinically relevant because it reflects early disease lethality, limited physiological reserve, and the period in which first-line treatment tolerance and timely palliative discussions are often determined. Existing prediction tools have largely addressed long-term survival, postoperative disease, or metastatic gastric cancer as a broad entity. This study developed and temporally validated an interpretable prediction model for six-month CSM in de novo metastatic GEJA.

METHODS: Patients diagnosed in 2010-2021 were identified from the Surveillance, Epidemiology, and End Results (SEER) database. Eligible cases had C16.0 cardia primary site, adenocarcinoma-related histology, malignant behavior, positive histology, single or first primary malignancy, de novo metastatic disease, and evaluable survival and cause-of-death information. Candidate predictors included demographic characteristics, histology, grade, tumor size, organ-specific metastases, diagnosis year, and first-course treatment-status fields. Patients diagnosed in 2010-2017 formed the development cohort; those diagnosed in 2018-2021 formed the temporal validation cohort. The endpoint was SEER-defined cancer-specific death within six months after diagnosis. Candidate models included logistic regression, elastic net, random forest, extreme gradient boosting (XGBoost), gradient boosting machine, and multilayer perceptron. Discrimination, calibration, Brier score, decision curves, threshold-specific sensitivity and specificity, and SHapley Additive exPlanations (SHAP) were evaluated.

RESULTS: The analytic cohort included 6,564 patients, with 4,315 in development and 2,249 in temporal validation. Median age was 64 years, 5,189 patients (79.1%) were male, and liver metastasis was present in 3,249 patients (49.5%). Six-month CSM occurred in 1,941 patients (45.0%) in development and 902 patients (40.1%) in temporal validation. The selected XGBoost model achieved an area under the curve (AUC) of 0.808 [95% confidence interval (CI), 0.795-0.821] in development and 0.786 (95% CI, 0.767-0.806) in temporal validation. In temporal validation, the Brier score was 0.173 versus a null Brier score of 0.240, the calibration slope was 1.130, and the integrated calibration index was 0.026. At the high-risk threshold, sensitivity was 56.7%, specificity was 90.2%, positive predictive value was 79.5%, and negative predictive value was 75.7%. Observed six-month CSM was 20.2%, 32.1%, and 79.5% in low-, intermediate-, and high-risk groups, respectively.

CONCLUSIONS: A SEER-based prediction model was developed and temporally validated for six-month CSM in de novo metastatic GEJA. The model provided moderate discrimination, acceptable calibration, and clinically distinct population-level risk strata. Because validation was limited to a later SEER cohort, the model should be externally validated and locally recalibrated before clinical workflow use, and it must not be interpreted as evidence of treatment efficacy.