Background: Esophageal squamous cell carcinoma (ESCC) accounts for over 80% of esophageal cancer cases globally. Accurate preoperative prediction of lymph node metastasis (LNM) is critical for treatment decision-making, yet current clinical staging frequently underestimates nodal disease burden. Pathomics — the high-throughput quantitative analysis of histopathological images — offers an opportunity to capture tumor microenvironmental patterns invisible to routine microscopy. We aimed to develop and validate a pathomics-clinical integrated model for predicting LNM in resectable ESCC. Methods: A total of 477 consecutive ESCC patients who underwent radical esophagectomy were retrospectively enrolled. After excluding 61 patients with missing data, 416 patients were included in the final analysis (LNM-positive: 217, 52.2%). H&E-stained whole-slide images were acquired using a KFBIO KF-PRO-400-HI scanner at 20× magnification (pixel size 0.5 μm/px). Nuclear morphological features were extracted via QuPath hematoxylin optical density (OD)-based detection, yielding 17 pathomic features selected by LASSO-logistic regression. A combined clinical-pathomics nomogram was constructed and internally validated via 10-fold cross-validation (out-of-fold, OOF), with apparent (resubstitution) performance reported alongside to quantify optimism. Secondary analyses examined survival stratification within node-negative (pN0) patients and across whole-cohort risk groups. The final nomogram was additionally applied, without refitting, to an independent external cohort of 101 ESCC patients from a separate institution to assess transportability. Results: The combined nomogram achieved an out-of-fold (cross-validated) AUC of 0.821 (95% CI 0.779–0.858), compared with 0.791 for the clinical model and 0.762 for the pathomics-alone model. In multivariable analysis, the pathomics score remained an independent predictor of LNM (adjusted OR 2.47, 95% CI 1.80–3.39; P < 0.001), with significant incremental value over clinical variables alone (likelihood-ratio χ² = 35.5, P = 2.56e-09). Risk stratification by nomogram score tertiles showed a clear gradient of LNM prevalence: 15.8% (low-risk), 55.1% (intermediate-risk), and 85.6% (high-risk) (Cochran-Armitage trend P < 0.001). In the pN0 subgroup (n = 199), patients with high pathomics risk had significantly worse overall survival (3-year OS: 77.2% vs. 94.3%; log-rank P = 0.005) and progression-free survival (3-year PFS: 75.5% vs. 88.9%; log-rank P = 0.007). Externally, in an independent cohort of 101 patients from a separate institution (LNM+ = 25, 24.8%), the combined model achieved an AUC of 0.774 (95% CI 0.663–0.870), the clinical model 0.801, and the pathomics-alone model 0.685, confirming partial transportability of the clinical component while the pathomics signature showed limited generalizability across centers. Conclusions: The integration of quantitative pathomic features with conventional clinicopathological variables provides robust predictive power for LNM in ESCC and identifies a subset of node-negative patients at elevated risk of poor outcomes. These findings support the potential utility of pathomics as a complementary tool for preoperative risk stratification in ESCC. External validation in an independent cohort confirmed partial transportability of the clinical component but highlighted limited generalizability of the pathomics signature, underscoring the need for multi-center harmonization before clinical deployment.