Objective : to develop a preoperative model for predicting 90-day mortality to optimize treatment strategy in patients with resectable esophageal cancer and Siewert type I gastroesophageal junction cancer. Material and Methods . A retrospective cohort study included 225 patients with resectable esophageal cancer (n=179) and Siewert type I gastroesophageal junction cancer (n=46). A total of 79 preoperative variables were analyzed. Risk factors were selected using univariate and multivariate logistic regression. The dataset was randomly split into training and validation cohorts (70/30). The predictive model was developed using l1-regularized multivariable logistic regression and the TabNet neural network featuring a sparse attentive architecture that enables estimation of feature importance. Discriminative ability and calibration were assessed using AUCROC analysis and the Hosmer–Lemeshow test. Results. The in-hospital mortality, 30-day mortality and 90-day mortality rates were 8.9 % (n=20), 1.3 % (n=3), and 6.2 % (n=14), respectively. The most significant predictors of mortality included preoperative weight loss >10 %, grade 3–4 dysphagia, ECOG performance status ≥2, COPD, preoperative dyspnea, history of major thoracic surgery, TNM stage N2, tumor proximity to or invasion of the trachea/main bronchi, cerebrovascular disease history, diabetes mellitus, and complications during or after radiotherapy. The logistic regression model demonstrated an AUC-ROC of 0.86 (95 % Ci 0.79–0.92) in the training cohort and 0.77 (95 % CI 0.72–0.83) in the validation cohort. Calibration was supported by non-significant Hosmer–Lemeshow χ 2 tests in both cohorts: χ 2 =1.98, df=8 (p=0.98) and χ 2 =4.72, df=8 (p=0.78), respectively. The TabNet neural network achieved improved discrimination with AUC-ROC values of 0.95 (95 % CI 0.92–0.98) in the training cohort and 0.86 (95 % Ci 0.82–0.96) in the validation cohort; calibration results were χ 2 =8.38, df=8 (p=0.39) and χ 2 =9.46, df=8 (p=0.30), respectively. The model has been implemented as a software application and can be used in real-world clinical scenarios. Conclusion . In this study, a machine learning-based model for preoperative 90-day mortality prediction was developed. It demonstrated good validity and high discriminative performance. The findings support the promise of machine learning methods for clinical risk prediction in cancer surgery and highlight their potential for implementation in routine healthcare practice.