Development and validation of an explainable machine learning model for predicting acute kidney injury after robot-assisted partial nephrectomy: a retrospective multicenter study

Abstract

Objective

Robot-assisted partial nephrectomy (RAPN) is an established, minimally invasive technique to treat patients with renal masses. The incidence of acute kidney injury (AKI) after RAPN is high and is associated with poor prognosis. This study aims to develop and validate an interpretable machine-learning model based on clinical features for individualized risk assessment of RAPN-AKI.

Methods

We retrospectively reviewed 325 patients undergoing RAPN at the Third Medical Center of PLA General Hospital (May 2022–Oct 2023) as the training dataset, and 146 from the Fifth Medical Center of PLA General Hospital (Nov 2023–Dec 2024) for external validation. Models were constructed using Boruta-selected features and eight machine learning algorithms. Performance was assessed by the area under the receiver operating characteristic curve (AUC), F1-score, accuracy, precision, calibration, and decision curve analysis (DCA). Shapley additive explanations (SHAP) interpreted feature contributions.

Results

The incidence of AKI in internal training and external validation datasets was 24.6% and 26%, respectively. The Boruta algorithm identified duration of renal artery blockade, preoperative serum creatinine (Scr), gender, body mass index (BMI), and age as important features. Among the eight machine learning models, the Gradient Boosting Machine (GBM) model demonstrated the best and most stable predictive outcomes in the internal training dataset (AUC = 0.889) and external validation dataset (AUC = 0.779). Both the calibration curve and DCA indicated better calibration and greater net benefit. SHAP analysis revealed the contribution of important features in the following order: duration of renal artery blockade, Scr, BMI, age, and gender. Dependency plots showed that duration of renal artery blockade > 22 min, Scr > 80 µmol/L, BMI > 25 kg/m², age > 60 years, and male were significantly associated with an increased risk of AKI.

Conclusion

The GBM model exhibited strong predictive performance in both internal training dataset and external validation dataset and has the potential to assist clinicians to identify the high-risk patients early, enabling timely interventions that may reduce the incidence of RAPN-AKI and improving clinical outcomes. While, the interpretable machine learning model is currently applicable only to patients with low-risk or normal preoperative renal function.

Introduction

With the widespread clinical application of imaging techniques, small renal cell carcinomas (RCCs) have been identified at an increasing frequency. Robot-assisted partial nephrectomy (RAPN) has emerged as an alternative to laparoscopic partial nephrectomy (LPN) for removal of renal tumors. In comparison to LPN and open partial nephrectomy, RAPN offers several advantages, such as shortened operative length, decreased blood loss, increased postoperative renal function, and good oncological outcomes [1, 2]. However, the intraoperative clamping of the renal artery remains unavoidable, resulting in ischemic kidney injury. Previous studies have reported that the incidence of acute kidney injury (AKI) after RAPN can be as high as 20% or more. Nevertheless, postoperative AKI is often regarded as a self-limiting condition or merely a surgical complication [3, 4]. AKI is a syndrome characterized by a rapid decline in estimated glomerular filtration rate (eGFR), manifested clinically as a sustained rise in serum creatinine (Scr) and reduction in urine output and is associated with poor patient outcomes, including higher mortality, a longer length of stay, and higher adverse event rates. Compared to patients without AKI, those with postoperative AKI have higher in-hospital mortality and risk of progression to chronic kidney disease (CKD) [5–7]. Therefore, early identification and intervention are essential for the prevention of AKI [8]. Previous studies have identified several risk factors, including male gender, advanced age, high body mass index (BMI), hypertension, diabetes, active congestive heart failure, and renal ischemia/reperfusion (I/R) injury [9–11]. However, there remains no conclusive research defining the risk factors for postoperative RAPN-AKI. Currently, most studies evaluating AKI risk factors employ multivariate logistic regression based on statistical regression analysis, but such evaluation has many limitations in ranking the relative importance of variables and addressing the complexity of their interactions.

With the rapid advancement of artificial intelligence (AI), machine learning, as a branch of AI, enables constructing high-quality predictive models by learning from vast raw datasets rather than using hard-coded instructions and can be categorized into three types of learning: supervised, semi-supervised, and unsupervised [12]. Supervised learning depends on labeled data to find the optimal prediction model suitable for the data set, and then using this optimal prediction model to predict the unlabeled data. The performance of model depends directly on the dataset’s quality. Higher label accuracy leads to more precise models and predictions, particularly in regression and classification tasks. Common machine learning algorithms for supervised learning include linear regression, regression trees, k-nearest neighbors (KNN), eXtreme Gradient Boosting (XGBoost), adaptive boosting (AdaBoost), neural networks, naïve Bayes, decision trees, support vector machines (SVM), and logistic regression. Conversely, unsupervised learning algorithms try to find hidden structures in unlabeled data and is predominantly used for association analysis, clustering, and dimensionality reduction. Typical unsupervised learning algorithms include sparse autoencoders, principal component analysis (PCA), k-means clustering, density-based spatial clustering of applications with noise (DBSCAN), and expectation maximization algorithm (EM) [13–15]. Furthermore, the implementation of shapley additive explanations (SHAP) technology facilitates the quantification of each feature’s marginal contribution to the model’s predictions, thereby enhancing transparency and providing a comprehensive understanding of feature importance. This advancement transforms machine learning from a “black box” process into an algorithm characterized by increased transparency and interpretability [14, 16]. For example, Fan et al. utilized machine learning to develop a predictive model for forecasting the prognosis of sepsis-associated AKI patients, demonstrating that the XGBoost model exhibited superior predictive performance. By applying SHAP to interpret the importance of variables within XGBoost model, their methodology could assist clinicians to design precise treatment strategies [17]. Similarly, Tseng et al. developed a machine learning model to predict AKI following cardiac surgery, potentially assisting clinicians in optimizing postoperative treatment strategies and reducing complications [18]. However, the studies mentioned above predominantly focus on septic and critically ill patients, highlighting a research gap concerning the prediction of postoperative AKI in patients undergoing RAPN. Therefore, this study employed eight machine learning algorithms to develop predictive models, specifically logistic regression, SVM, gradient boosting machine (GBM), neural network, XGBoost, KNN, AdaBoost, and LightGBM. Additionally, SHAP was utilized to interpret these predictive models, aiming to enhance clinicians’ understanding and clinical application of AKI prediction models, with the goal of providing early personalized risk assessment for patients with RAPN-AKI.

Materials and methods

Patient selection

Ethical approval for this retrospective study was granted by the Medical Ethics Review Panel of both the Third and Fifth Medical Centers of the Chinese People’s Liberation Army General Hospital (approval number: KS2024-001, KY-2025-1-17-1, clinical trial number: not applicable). The study was conducted in accordance with the principles outlined in the Declaration of Helsinki. A total of 325 RAPN patients treated at our center between May 2022 and October 2023 were retrospectively collected to form the internal training dataset. Additionally, 146 RAPN patients treated at the Fifth Medical Center of the PLA General Hospital between November 2023 and December 2024 were included as the external validation dataset.

The inclusion criteria for the study were as follows: (1) patients undergoing RAPN under general anesthesia, with T1aN0M0 renal carcinoma (characterized by clinical tumor size < 4 cm, R.E.N.A.L. score ≤ 6, absence of vascular invasion or metastasis, and tumors located away from the renal hilum and collecting system), (2) age > 18, (3) patients classified as American Society of Anesthesiologists (ASA) grades I-Ⅲ, (4) patients with normal preoperative cardiac, liver, and renal function. The exclusion criteria included: (1) patients with a solitary kidney, (2) patients with eGFR < 60 ml/min, (3) patients who experienced significant complications during or following the surgical procedure, (4) any intraoperative modifications to the planned surgical approach, such as conversion to open surgery or radical nephrectomy.

All RAPN procedures at each hospital were performed by a dedicated, experienced urological surgical team using the Da Vinci Xi Surgical System.

Anesthesia

Upon arrival in the operating room, patients underwent standard monitoring for electrocardiogram, noninvasive arterial blood pressure, and peripheral oxygen saturation. Standardized induction protocols were employed, which included the administration of midazolam at 0.03 mg/kg, etomidate at 0.2 mg/kg, propofol at 1.0 mg/kg, sufentanil at 0.3 µg/kg, and cisatracurium at 0.2 mg/kg. After endotracheal intubation, volume-controlled respiration was employed with respiratory parameters adjusted to a fresh gas flow rate of 2 L/min, tidal volume of 6 ~ 8 ml/kg, respiratory rate of 12 ~ 15 breaths/min, inspiratory/expiratory ratio of 1:2, and inhaled oxygen concentration of 60% to maintain arterial partial pressure of carbon dioxide (PaCO₂) within the range of 35 ~ 45 mmHg. All patients received intravenous-inhalation combined anesthesia involving inhalation of 2%~3% desflurane, intravenous administration of propofol at a rate of 4 ~ 12 mg/kg/h, and infusion of remifentanil at a rate of 5 ~ 15 µg/kg/h. The bispectral index (BIS) was maintained between 40 ~ 60. After induction, radial artery catheterization was performed for continuous arterial pressure monitoring. Intraoperative fluid management was primarily with crystalloids (Lactated ringer’s solution) and colloids (succinyl gelatin or hydroxyethyl starch 130/0.4) based on hemodynamic status. Blood pressure was maintained within ± 20% of the baseline value, and vasoactive agents (norepinephrine, phenylephrine, ephedrine, or urapidil) were administered as needed.

Data collection

Demographic and clinical variables including patient age, gender, ASA class, BMI, preoperative comorbidities (hypertension, diabetes, cardiovascular disease), Scr levels before surgery and at 48 h postoperatively, and eGFR calculated using the Cockcroft-Gault formula. For males: eGFR (ml/min) = (140 - age) × body weight (kg) / [0.818 × Scr (µmol/L)]; for females: eGFR (ml/min) = male eGFR × 0.85. Intraoperative details were also documented, including the volume of crystalloid infusion, volume of colloid infusion, duration of renal artery blockade, and operative duration.

Outcome definition

According to the diagnostic criteria of KDIGO guidelines: AKI was diagnosed when Scr increased ≥ 26.5 µmol/L within 48 h or Scr increased to 1.5 times of the baseline value within 7d, or when urine output was < 0.5 ml/kg/h for 6 h or more.

Data preprocessing

Patients who underwent RAPN at the Third Medical Center of the PLA General Hospital were included as the internal training dataset, while those from the Fifth Medical Center of the PLA General Hospital were used as the external validation dataset. Data were extracted from electronic medical records and anesthesia records. We used SPSS 22.0 software for data processing and analysis. Original data were cleaned by excluding patients with more than 30% missing values among the included indicators. For the remaining missing data, multiple imputation was applied with five imputations to handle the missing values. Multicollinearity was assessed, and features with a variance inflation factor (VIF) exceeding 5 were removed. Normally distributed measurement data are expressed as mean ± standard deviation (𝑥̅ ± SD), while non-normally distributed measurement data are expressed as median and interquartile range [M (P25, P75)]. Depending on whether the measurement data followed a normal distribution, independent samples t-test or Mann-Whitney U test was applied. Categorical variables were label-encoded and expressed as frequencies and percentages, and compared using the chi-square test or Fisher’s exact test, as appropriate.

Feature selection

Feature selection was conducted using the “Boruta” package in R software (version 4.3.2). The Boruta algorithm is a feature selection method based on random forests, creates a corresponding shadow variable for each original variable. By using random forests to calculate the importance scores of both original and shadow features, and through 50 iterations, features were compared with their shadow counterparts to identify all relevant features associated with RAPN-AKI. Features whose importance scores exceeded those of their shadow variables were considered “important,” while those with lower importance scores were deemed “unimportant” and subsequently excluded.

Model construction and evaluation

Based on the features selected by Boruta, prediction models were constructed using the “tidymodels” package. A total of eight algorithms—logistic regression, SVM, GBM, neural network, XGBoost, KNN, Adaboost, and LightGBM—were used to develop prediction models. Each algorithm underwent 5 repetitions of 10-fold cross-validation to ensure model stability. Grid search was used to determine the optimal hyperparameter values for each algorithm, maximizing the area under the receiver operating characteristic curve (AUC). Models were trained on the internal training dataset and validated on the external validation dataset. Model performance was evaluated using AUC, sensitivity (recall), specificity, accuracy, precision, and F1-score. The model with the highest AUC was considered the optimal model. Calibration curves and decision curve analysis (DCA) were plotted. The calibration curve assessed the reliability of the model’s predicted probabilities, with curves closer to the diagonal indicating better accuracy. The DCA curve evaluated the clinical utility of the model, where a higher threshold indicated greater net benefit.

Model interpretation

Global and local interpretations of the prediction model were conducted using the “kernelshap” and “shapviz” packages. For global interpretation, the Shapley algorithm was used to quantify the contribution of each feature to the prediction model. A positive SHAP value indicated that the feature increased the risk of AKI, while a negative value indicated a decreased risk. A beeswarm plot was used to depict the distribution and SHAP values of important features. A feature importance plot ranked features by their importance. Dependency plots were used to visualize the relationship between each important feature and postoperative AKI. For local interpretation, waterfall plots and force plots were employed to explain the specific contributions of each feature to the prediction output for individual samples.

A p-value < 0.05 was considered statistically significant.

Results

Patient demographic and clinical characteristics

We retrospectively enrolled 325 patients who underwent RAPN at our center between May 2022 and October 2023 as the internal training dataset, and additionally collected 146 patients who underwent RAPN at the Fifth Medical Center of the Chinese PLA General Hospital from November 2023 to December 2024 as the external validation dataset (Fig. 1). In the training dataset, 80 (24.6%) patients developed AKI, while 38 (26%) cases of AKI were observed in the validation dataset. No significant differences were found in baseline characteristics between the training and validation dataset (P > 0.05). To ensure model robustness, we performed multicollinearity diagnostics on the features. All VIF values were below 5, indicating no significant multicollinearity among the predictor variables (Table 1).

Fig. 1.

Flow chart of the study design. SVM: support vector machines, GBM: Gradient Boosting Machine, XGBoost: eXtreme Gradient Boosting, KNN: K-nearest neighbors, AdaBoost: Adaptive Boosting, SHAP: Shapley additive explanations

Table 1.

Clinical characteristics of patients with RAPN

Characteristics	ALL (n = 471)	Training dataset (n = 325)	Validation dataset (n = 146)	P
AKI [n (%)]	118 (25.1%)	80 (24.6%)	38 (26%)	0.74
Age (years)	54 (45, 62)	54 (44, 62)	56 (48.8, 63)	0.20
ASA [n (%)]				0.85
I	71 (15.1)	48 (14.8%)	23 (15.8%)
II	370 (78.6%)	255 (78.5%)	115 (78.8%)
III	30 (6.4%)	22 (6.8%)	8 (5.5%)
Gender (male)	307 (65.2%)	220 (67.7%)	87 (59.6%)	0.09
BMI (kg/m2)	25.2 (23.1, 27.6)	25.3 (23.1, 27.7)	24.9 (22.9, 27.2)	0.33
Hypertension [n (%)]	123 (26.1%)	78 (24%)	45 (30.8%)	0.12
Cardiovascular disease [n (%)]	32 (6.8%)	20 (6.2%)	12 (8.2%)	0.41
Diabetes [n (%)]	59 (12.5%)	37 (11.4%)	22 (15.1%)	0.26
Crystalloid solution (ml)	1500 (1000, 1500)	1500 (1500, 2000)	1500 (1500, 2000)	0.33
Succinyl gelatin injection (ml)	0 (0, 0)	0 (0, 0)	0 (0, 0)	0.95
Hydroxyethyl starch 130/0.4 NaCl (ml)	0 (0, 500)	0 (0, 500)	0 (0, 500)	0.36
Blood volume (ml)	50 (30, 100)	50 (30, 100)	50 (50, 100)	0.17
Urine output (ml)	200 (100, 300)	150 (100, 237.5)	200 (100, 300)	0.10
Duration of renal artery blockade (min)	20 (15, 26)	21 (16, 27)	20 (15, 24)	0.13
Surgery time (min)	140 (115,167)	135 (113.5, 165)	144.5 (118, 170.2)	0.11
Scr (µmol/L)	66 (57, 77)	67 (57, 77,5)	65 (54.8, 75)	0.23
eGFR (ml/min)	100.8 (92, 110.3)	108.6 (92.3, 130.5)	106.1 (90, 129.9)	0.70

The skewed distribution data is presented as median (interquartile range). The enumeration data are presented as [n (%)]. AKI: acute kidney injury, ASA: American Society of Anesthesiologists, BMI: body mass index, Scr: serum creatinine, eGFR: estimated glomerular filtration rate

Features selection

The Boruta algorithm was applied to the internal training dataset for feature selection over 50 iterations. Important features were identified based on the algorithm’s results, and a ridge plot was generated to visualize the findings. In the plot, the gray area represents the minimum, mean, and maximum Z-score of the shadow features used as a reference. Blue indicates features with importance greater than the shadow features, yellow denotes features with importance close to the shadow features requiring further iteration, and red signifies features with importance lower than the shadow features. A total of five variables were selected, ranked in order of importance as follows: duration of renal artery blockade, Scr, gender, BMI, and age (Fig. 2).

Fig. 2.

Features selection based on Boruta. Blue: Important features; Yellow: Features to be determined; Red: Non-essential features; Gray: Shadow features. Scr: serum creatinine, BMI: body mass index, eGFR: estimated glomerular filtration rate, ASA: American Society of Anesthesiologists

Model construction and evaluation

We developed 8 models using an internal training dataset: Logistic Regression, SVM, GBM, KNN, Adaboost, Xgboost, Neural Network, and LightGBM. The AUC values for each model were as follows: Logistic Regression (AUC = 0.770, 95% CI: 0.710–0.829), SVM (AUC = 0.756, 95% CI: 0.696–0.816), GBM (AUC = 0.889, 95% CI: 0.849–0.928), Neural Network (AUC = 0.798, 95% CI: 0.739–0.857), Xgboost (AUC = 0.825, 95% CI: 0.773–0.877), KNN (AUC = 0.919, 95% CI: 0.890–0.947), Adaboost (AUC = 0.691, 95% CI: 0.631–0.750), and LightGBM (AUC = 0.905, 95% CI: 0.860–0.950).The Accuracy values were 0.702, 0.717, 0.834, 0.745, 0.692, 0.791, 0.766, and 0.858, respectively. Precision values were 0.441, 0.449, 0.638, 0.488, 0.438, 0.541, 0.524, and 0.677, while F1-score were 0.569, 0.535, 0.690, 0.595, 0.587, 0.702, 0.531, and 0.739 (Table 2).

Table 2.

Performance of the RAPN-AKI prediction model

Model	AUC (95%CI)	Accuracy	Sensitivity	Specificity	Precision	F1-score
Training dataset
Logistic	0.770 (0.710 − 0.829)	0.702	0.8	0.669	0.441	0.569
SVM	0.756 (0.696 − 0.816)	0.717	0.662	0.735	0.449	0.535
GBM	0.889 (0.849 − 0.928)	0.834	0.75	0.861	0.638	0.69
Neural Network	0.798 (0.739 − 0.857)	0.745	0.762	0.739	0.488	0.595
Xgboost	0.825 (0.773 − 0.877)	0.692	0.887	0.629	0.438	0.587
KNN	0.919 (0.890 − 0.947)	0.791	1	0.722	0.541	0.702
Adaboost	0.691 (0.631 − 0.750)	0.766	0.537	0.841	0.524	0.531
LightGBM	0.905 (0.860 − 0.950)	0.858	0.812	0.873	0.677	0.739
Validation dataset
Logistic	0.765 (0.677 − 0.852)	0.685	0.816	0.639	0.443	0.574
SVM	0.762 (0.677 − 0.847)	0.644	0.816	0.583	0.408	0.544
GBM	0.779 (0.693 − 0.865)	0.705	0.789	0.676	0.462	0.583
Neural Network	0.691 (0.596 − 0.786)	0.658	0.658	0.657	0.403	0.5
Xgboost	0.734 (0.637 − 0.831)	0.678	0.763	0.648	0.433	0.552
KNN	0.748 (0.657 − 0.838)	0.747	0.579	0.806	0.512	0.543
Adaboost	0.581 (0.501 − 0.661)	0.726	0.289	0.88	0.458	0.355
LightGBM	0.675 (0.568 − 0.781)	0.705	0.579	0.75	0.449	0.506

SVM: support vector machines, GBM: Gradient Boosting Machine, XGBoost: eXtreme Gradient Boosting, KNN: K-nearest neighbors, AdaBoost: Adaptive Boosting

Subsequently, we validated the predictive performance of the models using an external validation dataset. The AUC values for each model were as follows (Fig. 3B): Logistic (AUC = 0.765, 95%CI: 0.677 − 0.852), SVM (AUC = 0.762, 95%CI: 0.677 − 0.847), GBM (AUC = 0.779, 95%CI: 0.693 − 0.865), NeuralNetwork (AUC = 0.691, 95%CI: 0.596 − 0.786), Xgboost (AUC = 0.734, 95%CI: 0.637 − 0.831), KNN (AUC = 0.748, 95%CI: 0.657 − 0.838), Adaboost (AUC = 0.581, 95%CI: 0.501 − 0.661), LightGBM (AUC = 0.675, 95%CI: 0.568 − 0.781). The accuracy values were 0.685, 0.644, 0.705, 0.658, 0.678, 0.747, 0.726, and 0.705. The precision values were 0.443, 0.408, 0.462, 0.403, 0.433, 0.512, 0.458, and 0.449. The F1-score were 0.574, 0.544, 0.583, 0.500, 0.552, 0.543, 0.355, and 0.506 (Table 2).

Fig. 3.

Machine learning models for predicting RAPN-AKI. A: ROC curves of the machine learning models in the internal training cohort; B: ROC curves of the machine learning models in the external validation cohort. Higher AUC indicates superior predictive performance. SVM: support vector machines, GBM: Gradient Boosting Machine, XGBoost: eXtreme Gradient Boosting, KNN: K-nearest neighbors, AdaBoost: Adaptive Boosting

We compared the AUC, sensitivity (recall), specificity, accuracy, precision, and F1-score of various models on both the internal training dataset and the external validation dataset. While the KNN and LightGBM models also exhibited strong predictive performance on the internal training set, the KNN model achieved a sensitivity of 1, indicating an excessively high false positive rate. Moreover, both the KNN and LightGBM models had lower AUC values on the external validation set compared to the GBM model, suggesting inferior predictive performance relative to the GBM model. Additionally, the AUC values of the KNN and LightGBM models dropped significantly in the external validation dataset, indicating severe overfitting. Therefore, we ultimately selected the GBM model as the optimal model. On the internal training dataset, the GBM model correctly identified 60 true positives and 211 true negatives, misidentified 34 false positives and 20 false negatives, with a true positive rate (sensitivity) of 75% and a true negative rate (specificity) of 86.1% (Fig. 4A). On the external validation dataset, it correctly identified 30 true positives and 73 true negatives, while incorrectly identifying 35 false positives and 8 false negatives, resulting in a true positive rate of 78.9% and a true negative rate of 67.6% (Fig. 4B).

Fig. 4.

Confusion matrix of the GBM model for predicting RAPN-AKI. A: Confusion matrix for the internal training dataset; B: Confusion matrix for the external validation dataset

Additionally, the calibration curves of the GBM model closely followed the ideal diagonal in both the internal training dataset and the external validation dataset, indicating that the predicted probabilities of postoperative AKI in RAPN patients align well with the actual probabilities (Fig. 5). The decision curve analysis for the internal training dataset showed that when the threshold probability was between 0.02 and 0.98, the net benefit provided by the GBM model was significantly higher than that of other models and the baseline (Fig. 6A). Similarly, on the external validation dataset, the model demonstrated favorable net benefits, particularly within the threshold probability range of 0.1 to 0.78, where it maintained a high level of net benefit (Fig. 6B).

Fig. 5.

Prediction of RAPN-AKI by machine learning models. A: Calibration curve of the machine learning model on the internal training dataset; B: Calibration curve of the machine learning model on the external validation dataset. The closer the calibration curve is to the diagonal line, the better the calibration performance. SVM: support vector machines, GBM: Gradient Boosting Machine, XGBoost: eXtreme Gradient Boosting, KNN: K-nearest neighbors, AdaBoost: Adaptive Boosting

Fig. 6.

Prediction of RAPN-AKI by machine learning models. A: DCA of machine learning models in the internal training dataset; B: DCA of machine learning models in the external validation dataset. A higher net benefit indicates better clinical utility of the model. SVM: support vector machines, GBM: Gradient Boosting Machine, XGBoost: eXtreme Gradient Boosting, KNN: K-nearest neighbors, AdaBoost: Adaptive Boosting

Model explanation

We employed SHAP analysis to quantify the influence of different features on the predictions of the GBM model and to evaluate the importance and contribution of each feature. As illustrated in Fig. 7A, the magnitude and distribution of SHAP values for each sample across different features can be observed. Yellow represents high feature levels, while purple indicates low levels. Points closer to the right suggest a stronger positive influence of the feature on the model output, whereas points closer to the left indicate a stronger negative influence (e.g., as the duration of renal artery blockade increases, its positive impact on predicting AKI also increases). Figure 7B displays the ranking of feature importance on the vertical axis, with the horizontal axis representing the mean SHAP value. Longer bars indicate a greater contribution of the feature to the model’s predictions. After SHAP interpretation, the ranking of key features by contribution is as follows: duration of renal artery blockade, Scr, BMI, age, and gender, revealing the true contribution of each feature on the predictive model.

Fig. 7.

Global model interpretation using SHAP. A: SHAP bee swarm plot, where each point represents the SHAP value of a feature for an individual patient. The color of the points indicates the feature value, with yellow corresponding to higher values and purple to lower values. B: SHAP bar plot, where the vertical axis displays the features ranked by importance, and the horizontal axis represents the mean SHAP value. Scr: serum creatinine, BMI: body mass index

To visualize the relationship between key features and postoperative AKI, we generated SHAP dependency plots for important features (Fig. 8). The horizontal axis represents the feature value, and the vertical axis shows the SHAP value. SHAP values above the zero threshold indicate a promoting effect of the feature on AKI prediction. Specifically, duration of renal artery blockade > 22 min, Scr > 80 µmol/L, BMI > 25 kg/m², age > 60 years, and male gender were significantly associated with an increased risk of AKI.

Fig. 8.

Dependency plot of SHAP. A: Duration of renal artery blockade; B: Scr; C: BMI; D: Age; E: Gender. The dependency plot demonstrates how each feature influences the output of the predictive model, with each point representing an individual patient. Scr: serum creatinine, BMI: body mass index

To enhance the interpretation of the model’s decision-making process at the individual level, we constructed SHAP waterfall plots (Fig. 9A), which illustrate the contribution and ranking of each feature to the model output for a single sample. Yellow bars represent positive contributions, while purple bars indicate negative contributions. The features were ranked in descending order of contribution as follows: duration of renal artery blockade, Scr, BMI, gender, and age. The baseline prediction value was 0.116, and after accumulating the contributions of all features, the final prediction value reached 1, indicating a high risk of postoperative AKI in the patient. Additionally, force plot interpretations were generated for two representative samples, demonstrating how each feature contributed to the final prediction value starting from the model’s baseline prediction. As shown in Fig. 9B, the sample with a positive prediction (male, BMI: 32.9, Scr: 90, duration of renal artery blockade: 42, age: 37) is displayed, while Fig. 9C illustrates the sample with a negative prediction (female, age: 58, BMI: 21.5, Scr: 64, duration of renal artery blockade: 26).

Fig. 9.

Local interpretation of the SHAP model. Yellow bars represent positive contributions and purple bars represent negative contributions. A: SHAP waterfall plot for an individual AKI patient in the internal training dataset; B: SHAP force plot for an individual AKI patient in the internal training dataset; C: SHAP force plot for an individual non-AKI patient in the internal training dataset. E[f(x)] denotes the baseline predicted value, while f(x) represents the final predicted value. Scr: serum creatinine, BMI: body mass index

Discussion

AKI is a common complication following RAPN and is closely associated with adverse outcomes. Currently, the diagnosis of AKI based on Scr and urine output has limitations, which may delay the timely recognition of AKI and increase the likelihood of false-negative rate. Therefore, early and precise diagnosis and identification of AKI are crucial for improving patient prognosis, enhancing perioperative safety, and promoting rapid recovery. In this study, an initial cohort of 325 patients undergoing RAPN was enrolled, and the Boruta algorithm was utilized to identify relevant features associated with RAPN-AKI. Based on these feature variables, eight different machine learning models were developed to predict the occurrence of RAPN-AKI. Among these models, the GBM model exhibited superior predictive performance, achieving an AUC of 0.889 and demonstrating favorable net benefits. The model subsequently underwent validation with an external cohort of 146 patients, yielding an AUC of 0.779 and demonstrating similarly favorable net benefits. Furthermore, the SHAP algorithm was utilized to interpret the contribution and importance of each feature variable in the GBM model for postoperative AKI. This methodology assists clinicians in reducing the “black box” problem by offering a clearer understanding of the impact of each feature, and thereby enabling accurate prediction of postoperative AKI.

Previous studies on predictive models for AKI have predominantly employed LASSO and logistic regression for feature selection, focusing solely on preoperative data while overlooking the impact of intraoperative anesthetic factors on postoperative AKI. To address this limitation, our study included both baseline preoperative and intraoperative parameters employing the Boruta algorithm for feature selection [3]. The Boruta algorithm, a feature selection technique based on random forests, functions by randomly shuffling each feature in the original dataset to create corresponding shadow features. The original and shadow features are then combined to construct a random forest model, wherein importance scores for both sets of features are calculated. Through multiple iterations, each original feature’s importance score is compared with the maximum importance score among the shadow features, identifying original features with higher importance scores as “important features.” Compared to other feature selection methods, Boruta provides higher stability, effectively reduces randomness and uncertainty in the feature selection process, and is capable of identifying all features relevant to AKI [19, 20]. Using Boruta, we identified five important features: duration of renal artery blockade, Scr, gender, BMI, and age. Although these risk factors have been extensively studied in previous research, our study integrates these variables into a clinical prediction tool to enable early prediction of RAPN-AKI. Consequently, our objective is not to discover new risk factors but to develop and validate an interpretable machine learning model based on common clinical characteristics.

Subsequently, machine learning models were constructed using the variables selected by Boruta. Unlike traditional logistic regression, machine learning techniques does not require the data to conform to statistical assumptions such as the independence of observations or the avoidance of multicollinearity among independent variables. These models have demonstrated strong performance in predicting outcomes for patients with AKI in the intensive care unit (ICU) [21, 22]. This study is the first application of machine learning to predict RAPN-AKI. By comparing eight machine learning models, we found that the KNN model achieved an AUC of 0.919, but with a sensitivity of 1 and a specificity of 72.2%, indicating a relatively high rate of false positives. In the external validation dataset, both KNN and LightGBM models produced lower AUC values compared to the GBM model, suggesting inferior predictive performance relative to GBM. Furthermore, the sharp decline in AUC for the KNN and LightGBM models from the training dataset (0.919, 0.905) to the validation dataset (0.748, 0.675) suggests severe overfitting. This discrepancy may be attributed to the limited sample size of the external validation dataset. Therefore, the GBM model was selected as the predictive model for postoperative AKI, as it demonstrated favorable net benefits and high threshold probabilities in both the internal training dataset and external validation dataset. The GBM model is a powerful machine learning technique for building non-parametric regression or classification models. It progressively and sequentially transforms weak models (e.g., decision trees) into strong ones. With each boosting iteration, each new model fits the negative gradient of the loss function from the previously aggregated models, thereby reducing the loss and achieving higher accuracy compared to other models [23]. Jiang et al. demonstrated that the GBM model exhibited excellent predictive capability for the occurrence of sepsis-associated persistent AKI, surpass in other models such as KNN, logistic regression, and Catboost [24]. Previous studies have also indicated that the GBM model exhibits optimal predictive performance for AKI following cardiac surgery, liver transplantation, and severe acute pancreatitis [25–27]. However, these studies predominantly focused on critically ill surgical patients and overlooked urological surgeries, which have a high incidence of AKI. Therefore, constructing a predictive model for AKI after RAPN is of significant importance.

With the increasing number of clinical studies focusing on machine learning, the lack of interpretability of input features has turned machine learning into an opaque black box, thereby constraining their clinical applicability [28]. Therefore, we employed SHAP to provide a global interpretation of the predictive model’s features. The global visualization revealed the following ranking of feature importance: duration of renal artery blockade, Scr, BMI, age, and gender. Subsequently, specific data were employed to locally interpret the prediction outcomes for individual patients undergoing RAPN, thereby assisting clinicians in better understanding the contribution of each feature and enhancing the model’s practical application in clinical settings.

We employed the Boruta algorithm, which identified age, gender, and BMI as significant features in predicting postoperative AKI, with their contributions ranked as following order: BMI, age, and gender. Notably, obese patients exhibit a higher incidence of mesangial matrix expansion, mesangial cell proliferation, podocyte hypertrophy, and glomerular hypertrophy compared to their non-obese patients. Additionally, they also frequently suffer from diabetes, hypertension, and hyperlipidemia. These metabolic abnormalities may promote the occurrence of AKI [29]. Xi et al. found that abdominal obesity might independently contribute to AKI in trauma patients by increasing intra-abdominal pressure, which resulted in renal vein compression and reduced renal perfusion [30]. Ahn et al. demonstrated that for every 10-unit increase in BMI, the risk of sepsis-associated AKI increases by 75% [31]. Wei et al. showed that a machine learning model constructed using features such as BMI can effectively predict the risk of AKI after cardiac surgery [32]. In our study, we found that a BMI exceeding 25 kg/m² was associated with an increased risk of AKI, although the specific pathophysiological mechanisms require further investigation. Previous studies have shown that elderly patients undergo a reduction in renal mass, a decrease in the number of functional glomeruli, and an increased rate of intrarenal apoptosis. These changes are accompanied by compensatory glomerular hypertrophy, atherosclerosis, interstitial fibrosis, fibrointimal hyperplasia, and hyalinization, leading to decreased cell proliferation, weakened self-renewal capacity, and a diminished regenerative response after kidney injury, thereby promoting the development of AKI [33, 34]. In predictive models for AKI among non-cardiac surgical patients, age over 60 has been identified as a significant risk factor and a critical preoperative variable, which is consistent with our findings [35]. Similarly, we found that patients over 60 years old were associated with a higher risk of AKI, however, its contribution was less than that of BMI. The SHAP dependence plot from our analysis indicated a significant association between male gender and an elevated risk of AKI. Prior research indicates that female animals have higher levels of renal endothelial nitric oxide synthase compared to males. Testosterone has been shown to reduces the activation of nitric oxide synthase induced by ischemia, while nitric oxide counteracts the pressor effects of angiotensin II, thereby reducing renal injury. Long-term estrogen therapy has been shown to reduce levels of angiotensin-converting enzyme and angiotensin II while increasing angiotensin 1–7, ultimately decreasing the incidence of AKI [36, 37]. Previous studies have reported that the incidence of AKI is higher in male patients across different age groups, and male AKI patients are more likely to progress to chronic kidney disease [38]. The SHAP results in our study indicated a relatively lower contribution of gender, which may be related to the relatively small number of female patients included in our analysis.

In our study, the duration of renal artery blockade emerged as the most critical predictor of AKI, as evidenced by global SHAP analysis, which highlighted its substantial contribution. The underlying mechanism may involve ischemia-induced imbalance between local tissue oxygen supply and demand, along with accumulation of metabolic waste products. These processes lead to injury of renal tubular epithelial cells, loss of brush borders, and eventual detachment of necrotic cellular debris into the tubular lumen, resulting in tubular dilation. Furthermore, tubular cast formation may occur as a result of necrosis and apoptosis [39]. Studies have shown that every 5-minute increment in renal ischemia time elevates the risk of postoperative AKI by 16%, while ischemia durations exceeding 25 min increases the risk of renal impairment by 6.5-fold compared to durations under 25 min. Therefore, it is recommended to establish a 25-minute threshold as the safe ischemic limit to reduce the incidence of kidney injury [40–42]. Our study’s SHAP dependence plot revealed that duration of renal artery blockade exceeding 22 min was associated with increased incidence of AKI. Thus, we recommend controlling the duration of renal artery blockade within 22 min during RAPN to minimize the risk of AKI.

Our investigation revealed a significant association between preoperative Scr levels exceeding 80 µmol/L and the incidence of AKI. As an indicator of renal function, elevated Scr implies substantial kidney injury, with perioperative stress further exacerbating AKI development [43]. Previous studies have revealed that higher preoperative Scr levels are significantly correlated with AKI incidence in heart transplant patients [44]. In a machine learning model designed to predict cardiac surgery-associated AKI, Li et al. identified preoperative Scr as an important risk factor, showing a positive correlation with postoperative AKI [45]. However, Scr levels are subject to variation due to factors such as age, race, gender, muscle mass, total body volume, medications, and protein intake. Therefore, incorporating additional variables is essential for the early diagnosis of AKI. In summary, we identified that advanced age, male gender, overweight, a preoperative Scr level >80 µmol/L, and the duration of renal artery blockade exceeding 22 min were associated with an increased risk of AKI. Thus, future studies could transform these characteristics into categorical variables with appropriate cut-off values for statistical analysis. Furthermore, the precise relationships between these variables and AKI warrant further investigation and discussion.

In our study, no collinearity was found among clinical characteristics, but key risk factors like hypertension and diabetes were not selected by Boruta, likely due to excluding patients with preoperative eGFR < 60 ml/min. Hypertension and diabetes are known to lower eGFR by damaging kidney function [46–48]. However, due to the small and mild comorbidity cases in our study, their impact on preoperative renal function was minimal, reducing statistical power. We also excluded patients with major intraoperative complications. Additionally, the included patients had shorter surgery times and minimal colloid infusion, which likely explains the lack of statistical significance for these factors. Consequently, weaker predictive risk factors were omitted from the final model.

For patients our model identified as high risk for AKI, a tailored perioperative management strategy can be applied. Preoperatively, surgeons can use this data to enhance surgical planning and inform patients about potential renal function changes, improving informed consent and follow-up planning. Intraoperatively, anesthesiologists can follow KDIGO guidelines to manage volume and hemodynamics. Our study shows that keeping renal artery clamp time under 22 min reduces AKI risk, aiding surgical decisions. Postoperatively, rigorous monitoring of Scr and urine output, along with avoiding nephrotoxic drugs, is crucial. This individualized approach can lower RAPN-AKI risk and enhance outcomes.

We acknowledge several limitations inherent in this study. Firstly, machine learning requires a large sample size to develop predictive models; however, our analysis was limited to one year of data from only two hospitals. This relatively limited sample size may reduce the model’s accuracy, and its applicability to a global population remains uncertain. Secondly, our prediction model relied on conventional clinical variables, critical variables such as intraoperative hemodynamic fluctuations, inflammatory markers, and early kidney injury biomarkers were not included. Moreover, the outcome assessment depended only on traditional renal function indicators like Scr, with a lack of postoperative urine output data. This reliance on Scr may lead to underdiagnosis of AKI and reduced predictive performance. Thirdly, over 99% of patients were discharged within five days after surgery, making it difficult to diagnose those with a rise in Scr occurring between 5 days and 7 days postoperatively and resulting in a lower detected incidence of AKI. Fourthly, due to the low incidence of AKI stages 2 and 3 in our study, we did not perform stratified statistical analysis by AKI stage, which may constrain the model’s capacity to accurately differentiate between various stages of AKI. Finally, the strict exclusion criteria applied in this study may restrict the model’s applicability to the highest risk patients for AKI. Future research should be directed toward expanding the sample size, relaxing exclusion criteria, incorporating early kidney injury biomarkers, extending postoperative follow-up duration and conducting prospective cohort validation to enhance the model’s generalizability and clinical utility.

Conclusion

The GBM model exhibited strong predictive performance in both internal training dataset and external validation dataset and has the potential to assist clinicians to identify the high-risk patients early, enabling timely interventions that may reduce the incidence of RAPN-AKI and improving clinical outcomes. While, the interpretable machine learning model is currently applicable only to patients with low-risk or normal preoperative renal function.

Abbreviations

AdaBoost

Adaptive Boosting

AI

Artificial intelligence

AKI

Acute kidney injury

ASA

American Society of Anesthesiologists

AUC

Area under the receiver operating characteristic curve

BMI

Body mass index

CKD

Chronic kidney disease

DCA

Decision curve analysis

eGFR

estimated glomerular filtration rate

GBM

Gradient Boosting Machine

KNN

K-nearest neighbors

LPN

Laparoscopic partial nephrectomy

OPN

Open partial nephrectomy

RAPN

Robot-assisted partial nephrectomy

RCC

Renal cell carcinoma

Scr

Serum creatinine

SHAP

Shapley additive explanations

SVM

Support vector machines

XGBoost

eXtreme Gradient Boosting

Author contributions

JXL designed the study and wrote the paper. YL, WWL, HG, FL, CHS, HXC, YFL and JH collected data. JXL, XFL, JY and LHX analyzed the data. YQY, LHX, YZL reviewed and edited the manuscript. YZL supervised all phases of the project. All the authors have read and approved the final manuscript.

Funding

The authors did not receive specific funding.

Data availability

All data supporting the results reported in this study are available from the corresponding author upon request.

Declarations

Ethics approval and consent to participate

The Medical Ethics Review Panel of the third Medical Center and fifth Medical Center of the Chinese People’s Liberation Army General Hospital granted approval to waive the requirement for informed consent for this study (approval number: KS2024-001, KY-2025-1-17-1, clinical trial number: not applicable). Its publication is also approved tacitly by the responsible authorities where the work was carried out. The study was conducted in accordance with the principles outlined in the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.