Jin Zhang, Lanyu Cai, Yunfei Kong, Zhongda Chen
The preoperative and combined models provide complementary, hypothesis-generating risk estimates at different clinical stages. They should be used only as adjuncts to established ESD criteria and multidisciplinary judgment. External validation in prospective, real-world ESD cohorts is required before clinical implementation.
AIM: To develop and temporally validate preoperative and combined preoperative-pathological models for lymph node metastasis (LNM) in early gastric cancer (EGC).
METHODS: We retrospectively analyzed 501 patients with EGC. Patients treated from January 2014 to December 2021 formed the training cohort (n = 347; 39 LNM events), and those treated from January 2022 to June 2026 formed the temporal validation cohort (n = 154; 21 LNM events). All preprocessing, exploratory LCR cut-off derivation, variable selection, coefficient estimation, and model choice were performed in the training cohort. The locked models were then evaluated without refitting in the temporal validation cohort. Performance was assessed using area under the receiver operating characteristic curve (AUC), Brier score, calibration intercept and slope, decision curve analysis, continuous net reclassification improvement (NRI), and integrated discrimination improvement (IDI). Conditional internal validation used 1,000 bootstrap samples with the final predictor sets held fixed.
RESULTS: The final preoperative model (PSL) included granulocyte count, albumin, LCR, tumor location, endoscopic ulceration, and long diameter. The combined model (Post-LR) included LCR, neurovascular invasion, long diameter, pathological ulceration, and differentiation. In temporal validation, the AUC was 0.730 (95% CI, 0.627-0.834) for PSL and 0.827 (95% CI, 0.723-0.932) for Post-LR. The corresponding Brier scores were 0.119 and 0.081. Compared with the pathology-only model, Post-LR had no statistically significant incremental benefit in temporal validation. An exploratory training-derived LCR cut-off of 2.20 separated risk in the training cohort but not in the temporal validation cohort.
CONCLUSION: The preoperative and combined models provide complementary, hypothesis-generating risk estimates at different clinical stages. They should be used only as adjuncts to established ESD criteria and multidisciplinary judgment. External validation in prospective, real-world ESD cohorts is required before clinical implementation.