Application of ML models to predict postoperative outcomes in renal cell carcinoma.

Abstract

463 Background: Accurately predicting postoperative outcomes, such as major and minor complications, readmissions, and mortality, enables personalized preoperative planning, informed decision-making, and improved perioperative management. In this study, we utilized a machine learning (ML) model to predict these outcomes in RCC patients undergoing nephroureterectomy, radical nephrectomy, partial nephrectomy, and other excision procedures on the kidney using data from the National Surgical Quality Improvement Program (NSQIP; 2016-2021). Methods: A gradient-boosted tree (GBT) model was developed and trained to predict the primary outcomes of interest: major complications, minor complications, 30-day readmission, and mortality. The NSQIP dataset provided a robust source of patient data, encompassing a wide array of variables, including demographic information, preoperative health status, comorbidities, intraoperative factors, and postoperative laboratory results. The model's performance was rigorously evaluated using various measures, such as AUROC, generalized R-square, MIS classification rate, etc. The model's predictions were validated using separate training (70%) and validation (30%) cohorts to ensure generalizability and applicability in diverse patient populations. Results: In the analysis, 36,284 cases were included. The GBT model demonstrated substantial predictive power across all four outcomes. In predicting mortality, the model achieved an AUROC of 0.98 in the training and 0.80 in the validation set, with a misclassification rate of 0.68%. The prediction of major complications was also robust, with AUROC of 0.94 in the training set and 0.91 in the validation set, accompanied by a misclassification rate of 2.2%. Similarly, the model's performance in predicting minor complications and 30-day readmissions was strong, with AUROC values of 0.70 and 0.66, respectively, in the validation sets. Among the most influential predictors identified by the model were age, diabetes, sepsis, bleeding disorders, ASA classification, and preoperative albumin levels. These factors consistently contributed to the model's accuracy across the different outcome measures. BMI was among the top 15 predictors for all four outcomes in GBT analysis. Conclusions: Our study demonstrated that using a national registry, ML models can accurately predict postop outcomes among patients with RCC. By integrating a wide range of patient-specific variables, the model offers a powerful tool for clinicians to identify high-risk patients and tailor perioperative care accordingly. These findings support broader ML adoption in surgical decision-making to enhance patient safety and optimize outcomes. Post-op complications in RCC by BMI category. BMI<30 BMI≥30 p-value Mortality 115 (0.64) 98 (0.54) 0.22 Major 580 (3.21) 607 (3.33) 0.49 Minor 798 (4.41) 998 (5.48) <0.0001 30-Day Readmission 1055 (5.83) 1045 (5.74) 0.70