Predicting 90-day postoperative mortality after esophagectomy using machine learning technology: a comparative evaluation of the model with foreign analogues

Objective. To compare the performance of a machine-learning-based model for predicting 90-day postoperative mortality after esophagectomy (EsoRisk) with existing prognostic scales, the ACS NSQIP and Esodata (IESG), in a single independent patient sample. Material and methods. A retrospective cohort study was conducted, including 100 patients with resectable esophageal and cardioesophageal junction cancer (Siewert I) who underwent esophagectomy at the P.Herzen Moscow Oncology Research Institute and the A.Tsyb Medical Radiological Research Center from January 2024 to November 2025. During the 90-day postoperative period, in-hospital, 30-, and 90-day mortality, as well as the incidence of postoperative complications, were assessed. The ACS NSQIP, Esodata (IESG), and EsoRisk scores were used for direct comparative analysis. Model validity was assessed using the AUC-ROC, AUPRC, Accuracy, and Brier score metrics. Results. Postoperative complications were registered in 42% of patients, Clavien–Dindo complications ≥ IIIA — in 29%. In-hospital mortality was 8%, 30-day — 2%, 90-day — 8%. In terms of discriminatory ability, EsoRisk was superior to the compared scales: AUC-ROC was 0.87 versus 0.78 for Esodata (IESG) and 0.55 for ACS NSQIP; AUPRC — 0.81 versus 0.36 and 0.36, respectively; Accuracy — 0.70 versus 0.54 and 0.46; the Brier score was minimal in the EsoRisk model and amounted to 0.10. Conclusion. The EsoRisk model, based on machine learning methods, demonstrated higher accuracy in predicting 90-day postoperative mortality after esophagectomy compared to the ACS NSQIP and Esodata (IESG) scores. These data confirm the potential of using machine learning algorithms for preoperative risk stratification in patients with esophageal cancer. The presented model can be recommended for use in institutions specializing in the surgical treatment of esophageal cancer.