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

Julia Mühlbauer, Luise Ingvelde Monika Gottstein, Luisa Egen, Caelán Max Haney, Alexander Studier‐Fischer, Evangelia Christodoulou, Giovanni Cacciamani, Keno März, Lena Maier‐Hein, Stephan Maurice Michel, Allison Quan, Karl‐Friedrich KowalewskiBJUI 2025 · DOI

TL;DR

This study explores the use of Artificial Intelligence (AI) and Machine Learning (ML) models to predict major complications (MCs) and acute kidney injuries (AKIs) after kidney cancer surgery. The findings suggest that Neural Networks (NNs) are the most effective models for this task, outperforming traditional logistic regression and other ML approaches, offering a promising avenue for personalized patient care.

What Problem Does This Paper Address?

Kidney cancer surgery, whether partial or radical nephrectomy, carries significant risks, including major complications and acute kidney injuries (AKIs), which can occur in up to 35% of patients. Accurate preoperative risk stratification is crucial for tailoring patient management and monitoring, but current tools are limited by small sample sizes, outdated data, and a lack of external validation. This paper addresses the unmet need for reliable, data-driven tools to predict these postoperative complications, aiming to improve patient outcomes and optimize surgical planning.

Methods

The researchers employed a comparative feasibility study design, evaluating several machine learning models against traditional logistic regression. The ML models investigated included:

  • Random Forest (RF): An ensemble learning method that constructs multiple decision trees.
  • Neural Networks (NN): Models inspired by the structure and function of the human brain, capable of learning complex patterns.
  • Support Vector Machines (SVMs): Algorithms that find an optimal hyperplane to separate data points.
  • Stochastic Gradient Boosting (GBM): An iterative process that builds an ensemble of weak learners.
  • Elastic Net Logistic Regression (ENET): A regularized regression method that combines L1 and L2 penalties, serving as a benchmark for traditional logistic regression.

Data preprocessing involved imputation, one-hot encoding, and handling of class imbalance through various sampling techniques. Model performance was optimized for sensitivity to minimize the risk of missing adverse events. The models were trained, tuned, and validated using a split dataset (75% training, 10% tuning, 15% validation). Performance was assessed using AUROC (Area Under the Receiver Operating Curve), sensitivity, Brier score (for calibration), and net benefit (for clinical utility).

Datasets

The study utilized a retrospective dataset of 963 patients who underwent partial nephrectomy (PN) or radical nephrectomy (RN) at the University Medical Center Mannheim between January 2017 and March 2023. This dataset comprised 30 variables, including 18 descriptive and 12 predictor variables. No new benchmarks were introduced; the study focused on applying and comparing existing ML models to this specific clinical dataset.

Key Insights and Ideas

  • NNs Show Superiority: Neural Networks emerged as the most promising models for predicting both major complications and AKIs, demonstrating better overall fit and sensitivity compared to other ML algorithms and logistic regression.
  • The Importance of Sensitivity: The study prioritized sensitivity in model optimization, reflecting the clinical imperative to avoid missing potential complications, even if it means slightly overestimating risk and leading to more cautious monitoring.
  • Predictor Variable Insights: For major complications, key predictors included open PN technique, pre-operative eGFR, and pre-operative hemoglobin. For AKI, tumor size, BMI, and the use of robot-assisted PN were identified as significant factors. This provides actionable insights for clinicians.
  • Class Imbalance Challenges: The study acknowledged the significant class imbalance for major complications (8.9% of patients) and how this can influence model performance and potentially lead to overfitting. The models for AKI, with a higher event rate (33.4%), showed better performance metrics, suggesting that class imbalance is a critical factor to consider in predictive modeling.

Key Results

  • Major Complications: The NN model achieved an AUROC of 0.762 [95%CI 0.611-0.912] and a sensitivity of 0.86 [95%CI 0.80-0.92]. While the SVM model had a higher AUROC (0.828), the NN was favored for its sensitivity. The ENET (logistic regression) model performed comparably to the NN in AUROC but was outperformed in sensitivity.
  • Acute Kidney Injury (AKI): The NN model achieved an AUROC of 0.717 [95%CI 0.611-0.823] and a sensitivity of 0.82 [95%CI 0.74-0.90]. The Random Forest model had a slightly higher AUROC (0.743), but the NN was noted for its strong net benefit. The ENET model performed poorly for AKI prediction, with the lowest sensitivity and highest Brier score among the models.

Limitations and Future Work

The authors acknowledge several limitations:

  • Cohort Size: The cohort of approximately 1000 patients, while substantial, is considered relatively small compared to multicentric studies, potentially limiting the ML models' ability to learn complex relationships. External validation on larger datasets is recommended.
  • Retrospective Data Collection: The retrospective nature of the data may have introduced biases or affected data quality and completeness.
  • Lack of Transparency: The "black box" nature of ML algorithms makes it difficult to fully comprehend the cause-effect relationships, which can hinder clinical adoption.
  • Electronic Data Collection: The lack of standardized, uniform electronic data collection across countries poses a hurdle for widespread ML application.

Future work should focus on external validation of these models, developing nomogram calculators for clinical application of the MC prediction model, and translating ML algorithms into more interpretable and user-friendly formats for everyday clinical practice.

Summary updated

Abstract

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.
BibTeX
@inproceedings{muhlbauer2025aidriven,
  author = {Julia Mühlbauer and Luise Ingvelde Monika Gottstein and Luisa Egen and Caelán Max Haney and Alexander Studier‐Fischer and Evangelia Christodoulou and Giovanni Cacciamani and Keno März and Lena Maier‐Hein and Stephan Maurice Michel and Allison Quan and Karl‐Friedrich Kowalewski},
  title = {{AI‐driven preoperative risk assessment in kidney cancer surgery: A comparative feasibility study of machine learning models}},
  booktitle = {BJUI},
  year = {2025},
  doi = {10.1002/bco2.70080},
}
Figure