A Novel Predictive Model for Cancer-Specific Surgical Risk Stratification of Patients Undergoing High-Risk Upper Gastrointestinal Oncologic Resections

Objective: High-risk oncologic resections are the primary curative interventions for solid upper gastrointestinal (GI) tumors; however, there are no cancer-specific risk-stratification models available for these operations. Methods: We studied 2823 Memorial Sloan Kettering Cancer Center patients (2015-2022) undergoing upper GI oncologic resections including gastrectomy, esophagectomy, distal pancreatectomy, and pancreaticoduodenectomy for adenocarcinoma. Univariable analyses assessed risk factors for length of stay (LOS) and complications, with multivariable analyses incorporating covariates significant on univariable analysis and procedure type. Nomograms were built using multivariable models to predict surgical complications and LOS. Model performance was evaluated using discrimination (C-index for complications, Kendall’s Tau for LOS) and calibration plots and was internally validated with 1000 bootstrap replicates to obtain average performance metrics and 95% confidence intervals. Results: The median LOS was 7.0 (IQR:6.0,10.0), and the overall 30-day postoperative complication rates were 31%. Multivariable analysis identified age, ASA>3, COPD, renal failure, and operation length as independent risk factors for increased LOS (OR>1, P <0.05). Neoadjuvant chemotherapy (NAC) and pre-incision antibiotics reduced LOS (OR<1, P <0.05). The complication nomogram demonstrated a C-index of 0.66, while the LOS nomogram showed a Kendall’s Tau of 0.39. Procedure type, age, and history of COPD and NAC had the greatest discriminatory values for LOS on the nomograms; procedure type, race, and BMI had the greatest impact on 30-day surgical complications. Conclusion: Our cancer-specific pre- and postoperative nomograms for LOS and postoperative complications provide an important tool to appropriately estimate risks and treatment outcomes for patients undergoing high-risk GI oncologic resections.