AI-driven preoperative risk assessment in kidney cancer surgery: A comparative feasibility study of machine learning models

Background and Objective

Preoperative risk stratification in renal tumour surgery is essential to enable risk-adjusted postoperative patient monitoring. Machine learning (ML) models predicting major complications (MCs) and acute kidney injuries (AKIs) following partial (PN) or radical nephrectomy (RN) have not been made, nor have they been compared with traditional logistic regression models.

Design, setting and participants

A total of 963 patients who underwent PN and RN between January 2017 and March 2023 at the University Medical Center Mannheim were included. The dataset consisted of 30 variables of interest– 18 descriptive and 12 predictor variables, which allowed for 7 predictor variables per event. The dataset was pre-processed, and ML models were created for MC and AKI. The selected models included Random Forest (RF), Support Vector Machines (SVMs), Stochastic Gradient Boosting, Neural Networks (NNs) and Elastic Net Logistic Regression models (ENETs).

Results and limitations

For major complications, the NN model had the best model fitting, with an AUROC of 0.762 [95%CI 0.611–0.912], a sensitivity of 0.86 [95%CI 0.80–0.92] and a Brier score of 0.17 [95%CI 0.11–0.23]. For AKI, the best fit model was created using a NN with an AUROC of 0.717 [95%CI 0.611–0.823], a sensitivity of 0.82 [95%CI 0.74–0.90] and a Brier score of 0.24 [95%CI 0.17–0.31]. The best performing models for both outcomes outperformed the ENETs.

Conclusions

The ML models provide valuable information for preoperative risk stratification of patients undergoing renal tumour surgery. This study suggests that NNs are the most appropriate models to stratify patients regarding the occurrence of MCs and AKIs, respectively. The models are made publicly available for reproducibility.