Risk Assessment for Upper Urinary Tract Deterioration in Patients With Neurogenic Bladder: A Retrospective Study Using Interpretable Machine Learning Algorithms
Article information
Abstract
Purpose
Upper urinary tract deterioration (UUTD) is a burdensome complication in patients with neurogenic bladder (NB). This study aimed to develop a practical machine learning (ML) tool for assessing UUTD risk.
Methods
The study included 2,393 patients with NB who underwent clinical and urodynamic examinations. Of these, 2,001 patients from northern regions of our country were assigned to training and validation sets, and 392 patients from southern regions formed the test set for internal-external validation. Clinical and urodynamic parameters were screened using the Boruta, least absolute shrinkage and selection operator regression, and decision tree algorithms; features identified by all 3 algorithms were used to develop the ML models. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis. SHapley Additive exPlanations analysis provided visual interpretations for individual patients.
Results
Six features, notably bladder management method and detrusor sphincter dyssynergia, were identified as risk factors for UUTD. Of the 4 ML algorithms, the Gaussian Naive Bayes (GNB) model had the best overall performance, with AUCs (95% confidence intervals) of 0.884 (0.867–0.901), 0.868 (0.829–0.907), and 0.878 (0.839–0.916) in the training, validation, and test sets, respectively. Subgroup analyses by imaging outcome, NB etiology, and age showed consistent performance of the GNB model, with AUCs ranging from 0.815 to 0.895. The GNB model incorporating the core variables was presented through a practical web-based tool for individualized risk assessment.
Conclusions
We developed a ML model to assess UUTD risk in patients with NB. The findings emphasize the importance of combining bladder management method with urodynamic evaluation when assessing UUTD risk, and the model provides a practical tool for individualized assessment. Independent external validation of the model is warranted.
INTRODUCTION
Neurogenic bladder (NB) is common among patients with neurological diseases, affecting approximately 83%–95% of those with spinal cord injury and 96% of those with spina bifida [1]. Impaired communication between the nervous system and the lower urinary tract characterizes NB and can lead to progressive upper urinary tract deterioration (UUTD). Despite advances in urological management, UUTD remains the main complication of NB, with 14.6% to 35.2% of patients developing varying degrees of renal impairment [2, 3].
Identifying UUTD early is critical to timely intervention and the prevention of irreversible renal damage. Previous studies have used traditional statistical methods to examine relationships between UUTD and urodynamic parameters [3], including detrusor leak point pressure (DLPP) [4-6] and bladder compliance [7, 8]. These studies have informed UUTD risk assessment, but their focus on individual variables or prespecified interactions limits their ability to capture complex interactions among multiple contributing factors. Furthermore, no large-scale study has comprehensively evaluated UUTD risk models across distinct NB subgroups, despite the diversity of the affected population described above.
Machine learning (ML) algorithms have emerged in recent years as powerful tools for analyzing complex data and identifying nonlinear relationships [9-11]. Given the complex effects of clinical and urodynamic parameters on UUTD [3], we aimed to develop an interpretable ML model for UUTD risk assessment using integrated clinical and urodynamic data from a large cohort of patients with NB. The study sought to provide clinicians with a practical tool for identifying patients at elevated risk of UUTD and supporting more informed, personalized management plans (Fig. 1). Reporting followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis+Artificial Intelligence statement [12] and the Transparency in the Reporting of Artificial Intelligence guideline [13].
MATERIALS AND METHODS
Study Participants
Clinical and urodynamic data were retrospectively collected from patients with NB at our center between January 2007 and August 2025. Patients who underwent video-urodynamics (VUDS) at their first visit to our center were included. Exclusion criteria were (1) urinary diversion sur-gery before the initial visit; (2) kidney dysfunction caused by hypertension, heart disease, renal diseases, or other systemic conditions; (3) ureteral dilation or hydronephrosis caused by urinary stones; (4) congenital malformations of the renal pelvis or ureter; and (5) incomplete clinical data that would affect subsequent analysis.
Patients were classified as being from northern or southern regions according to their household registration status and permanent residential address. The 2,001 eligible patients with NB from northern regions were randomly allocated to the training and validation sets in a 7:3 ratio. The 392 patients from southern regions formed the test set for internal-external validation (Fig. 1). The institutional review board at the China Rehabilitation Research Center approved this study (No. 2023-075-01), which was conducted in accordance with the Declaration of Helsinki.
Clinical Assessment
VUDS followed the “Good Urodynamic Practices” recommended by the International Continence Society (ICS) [14]. The following urodynamic data were collected in accordance with ICS guidelines: maximum cystometric capacity (MCC), volume at first sensation of bladder filling, volume at strong desire to void, maximum detrusor pressure (Pdetmax) during the filling phase, end-filling pressure, presence of detrusor overactivity (DO), bladder compliance (calculated manually from urodynamic traces), DLPP, detrusor sphincter dyssynergia (DSD), detrusor underactivity (DU), acontractile detrusor (AcD), and postvoid residual urine volume. DO was defined as any involuntary detrusor contraction during the filling phase. Bladder sensation was classified as normal, increased, reduced, or absent according to the corresponding criteria [15]. Because DLPP was not observed in all patients [7], end-filling pressure was used in subsequent analyses when DLPP was unavailable [16, 17]. Maximum urethral closure pressure (MUCP) was measured using urethral profilometry.
NB etiology was recorded and classified into 3 groups: spinal cord injury, spinal dysplasia, and other etiologies (including iatrogenic pelvic neuropathy, myelitis, diabetic bladder, and others). Bladder management methods used for more than 6 months were categorized as self-voiding, abdominal pressure-assisted urination, clean intermittent catheterization (CIC), indwelling catheterization, or other methods (such as use of an external collection device). Common upper urinary tract imaging examinations included cystourethrography, urinary tract ultrasonography, magnetic resonance urography (MRU), and radioisotope renography, performed concurrently at the initial clinical visit. All included patients underwent at least 2 of these imaging modalities to determine upper urinary tract status. Vesicoureteral reflux (VUR) was determined according to the International Reflux Grading System [18]. MRU and/or urinary tract ultrasonography was used to assess hydronephrosis [19-21], and radioisotope renography was used to determine measured glomerular filtration rate (mGFR). UUTD was defined as at least one of the following upper urinary tract abnormalities: VUR, hydronephrosis, or mGFR below 60 mL/min/1.73 m2.
Data Preprocessing
Multiple imputation was used for variables with a low proportion of missing values. For variables with a high proportion of missing values (≥30%), data missing important patient information were excluded when no reasonable substitute value was available. DLPP was missing in more than 30% of cases; given its potential value and relationship to end-filling pressure, the latter was substituted for missing DLPP values. Continuous variables were standardized using z-score normalization. Binary categorical variables were coded as 0 or 1, and unordered multicategory variables were converted to dummy variables using one-hot encoding. To avoid data leakage, all preprocessing steps were applied independently to the training, validation, and test sets.
Model Development
Feature selection comprised dataset analysis and feature prioritization. Following initial screening, feature importance in the training set was assessed using 3 algorithms: Boruta, least absolute shrinkage and selection operator (LASSO), and decision tree (DT). A Venn diagram [22] was used to identify features selected by all 3 algorithms for inclusion in the UUTD risk model.
Four widely used ML algorithms were used to develop the models: support vector machine (SVM), multilayer perceptron (MLP), Gaussian Naive Bayes (GNB), and k-nearest neighbor (KNN). To address potential class imbalance in upper urinary tract outcomes, the synthetic minority oversampling technique was applied exclusively to the training set before model training. Hyperparameters for each algorithm were tuned by grid search with 10-fold cross-validation in the training set; only the optimal nondefault settings are reported below. SVM uses supervised learning to minimize structural risk and improve generalization; the optimal settings in this study were “C=0.1, max_iter=50.” MLP stacks multiple layers of neurons with nonlinear activation functions to address linearly inseparable problems [23]; its optimal settings were “activation=logistic, hidden_layer_sizes=(30, 30), max_iter=100.” GNB handled categorical features as pre-encoded variables [9, 24], with “var_ smoothing=1e-07” to maintain computational efficiency and robustness to irrelevant features. KNN used distance metrics optimized for continuous variables, with pre-encoded categorical variables also included as input features; its final settings were “n_neighbors=6, leaf_size=10.”
Statistical Analysis and Model Evaluation
Baseline continuous variables were compared using 1-way analysis of variance or Student t-test, and categorical variables were compared using Fisher exact test or the chi-square test. Models were trained on the training set and evaluated on the validation and test sets. Performance was assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC), together with calibration, decision, and precision-recall curves. Subgroup analyses by upper urinary tract imaging outcome, NB etiology, and age were performed in the validation set to assess model performance across patient subgroups (Fig. 1). SHapley Additive exPlanations (SHAP) analysis was used to explain each feature’s contribution to the model. All statistical analyses were performed in R ver. 4.2.3 (R Foundation for Statistical Computing, Austria) and Python ver. 3.11.4 (Python Software Foundation, USA), with P-values <0.05 considered statistically significant.
RESULTS
Patient Characteristics
The training, validation, and test sets included 1,400, 601, and 392 patients with NB, respective-ly; Table 1 summarizes their characteristics. Across all patients, the proportions with VUR, hydronephrosis, and UUTD were 19.64%, 24.28%, and 32.97%, respectively (Table 1). Upper urinary tract imaging outcomes did not differ significantly among the training, validation, and test sets. Age, bladder management method, MCC, and DSD differed significantly among the 3 sets, whereas the other clinical and urodynamic variables did not (Table 1). Most clinical and urodynamic variables differed significantly between patients with and without UUTD; the exceptions were sex, DO, AcD, DU, and MUCP (Table 2).
Feature Selection
Features were selected in the training set using Boruta (Fig. 2A), LASSO regression (Fig. 2B and C), and DT (Fig. 2D). Venn diagram analysis identified 6 features common to all 3 methods: disease duration, abdominal pressure-assisted urination, bladder compliance, DSD, MCC, and Pdetmax (Fig. 2E).
Feature selection using multiple algorithms. (A) Important variables selected by the Boruta algorithm (green) according to z-scores. (B and C) LASSO regression: coefficient paths (B) and cross-validation curve (C). (D) The 10 most important variables selected by the DT algorithm according to coefficient values. (E) Venn diagram showing the overlap in features selected by the 3 algorithms. LASSO, least absolute shrinkage and selection operator; DT, decision tree.
Performance of ML Models
The 6 selected variables were used as independent factors to develop UUTD risk assessment models. Ten-fold cross-validation in the training set provided mean performance estimates for the 4 ML models (Table 3; Fig. 3). In the training set, KNN outperformed the other models, with an AUC of 0.932 and a 95% confidence interval (CI) of 0.921–0.943, compared with 0.884 (95% CI, 0.867–0.901) for GNB, 0.779 (95% CI, 0.754–0.804) for MLP, and 0.792 (95% CI, 0.770–0.815) for SVM (Fig. 3). In the validation and test sets, however, GNB outperformed the other models. Its AUCs in these sets were 0.868 (95% CI, 0.829– 0.907) and 0.878 (95% CI, 0.839–0.916), respectively. Fig. 3 compares the ROC, calibration, decision, and precision-recall curves for the 4 models. GNB had Brier scores of 0.131 and 0.134 in the validation and test sets, respectively, lower than those of the other models (Fig. 3D and G). Decision curve analysis (DCA) showed that GNB provided greater net clinical benefit than the other models across a broad range of thresholds (10%–90%) covering most clinical scenarios (Fig. 3E and H). Precision-recall curves likewise showed better performance for GNB than for the other models (Fig. 3F and I). Table 3 presents the remaining performance metrics for each model across the 3 sets.
Comparison of multiple machine learning evaluation indexes in the training, validation, and test sets
Performance of multiple machine learning algorithms in the training, validation, and test sets. (A–C) Comparison of ROC curves in the training (A), validation (B), and test (C) sets. (D) Calibration curves in the validation set. (E) Decision curve analysis in the validation set. (F) PR curves in the validation set. (G) Calibration curves in the test set. (H) Decision curve analysis in the test set. (I) PR curves in the test set. ROC, receiver operating characteristic; PR, precision-recall; GNB, Gaussian Naive Bayes; MLP, multilayer perceptron; SVM, support vector machine; KNN, K-nearest neighbor.
Subgroup analyses by upper urinary tract imaging outcome, NB etiology, and age were then performed in the validation set to assess ML model performance across patient subgroups (Table 4; Fig. 4). GNB outperformed the other models across the subgroups, with AUCs ranging from 0.815 to 0.895. For GNB, the AUCs for VUR, hydronephrosis, and the composite outcome (UUTD) were 0.815, 0.832, and 0.868, respectively, suggesting that the composite outcome might reflect upper urinary tract status more comprehensively than a single modality. Fig. 4 presents the ROC curves for GNB in each subgroup.
Subgroup analysis results of the validation set according to imaging outcome, neurogenic bladder etiology, and age
Subgroup analyses of the GNB model by imaging outcome, NB etiology, and age. (A–C) ROC curves for VUR (A), hydronephrosis (B), and UUTD (C). (D–F) ROC curves by NB etiology: spinal cord injury (D), spinal dysplasia (E), and other etiologies (F). (G and H) ROC curves by age: patients younger than 18 years (G) and aged 18 years or older (H). (I) Radar plot of the area under the ROC curve in each subgroup. GNB, Gaussian Naive Bayes; NB, neurogenic bladder; ROC, receiver operating characteristic; VUR, vesicoureteral reflux; UUTD, upper urinary tract deterioration.
Based on the ROC curves, calibration plots, DCA, precisionrecall curves, and subgroup analyses, GNB was selected as the final model for assessing UUTD risk in patients with NB. SHAP plots were used to visualize how the features affected GNB model output. Each feature’s contribution to model output was quantified and displayed in SHAP summary plots (Fig. 5A) to identify the most relevant risk factors. The SHAP importance plot (Fig. 5B) further illustrates the influence of individual features on GNB. In descending order of contribution, the features were abdominal pressure-assisted urination, DSD, disease duration, MCC, Pdetmax, and bladder compliance (Fig. 5A and B). A web-based calculator incorporating these 6 variables in the GNB model was developed to support individualized UUTD risk assessment and improve accessibility and ease of use (https://www.xsmartanalysis.com/model/list/predict/model/html?mid=31532&symbol=6177086jXjm8TO993Oo3; Fig. 5C). Fig. 5D shows SHAP force plots for 2 individual samples, with the calculated risk probabilities and corresponding explanations.
SHAP analysis of the GNB model for UUTD risk assessment. (A) SHAP summary plot. The horizontal position of each point represents its SHAP value, indicating a feature’s effect on an individual patient’s risk (red=high, blue=low). (B) SHAP importance ranking, with features ordered vertically by mean absolute SHAP values. (C) Calculator for UUTD risk assessment. (D) Individualized risk assessment for 2 samples. Horizontal plots show the magnitude of feature effects, and output values represent the probability of UUTD. These plots show variable importance and individual risk contributions within the GNB model. SHAP, SHapley Additive exPlanations; GNB, Gaussian Naive Bayes; UUTD, upper urinary tract deterioration; Pdetmax, maximum detrusor pressure; MCC, maximum cystometric capacity; DSD, detrusor sphincter dyssynergia.
DISCUSSION
Previous studies have examined several factors associated with UUTD risk, although their clinical value, applicability, and ease of use remain limited to some extent [7, 8, 25]. Comprehensive models are therefore still needed to assess UUTD risk in patients with NB. We developed an ML model for this purpose using a large cohort of patients with NB. Data-driven feature selection identified bladder management method, DSD, disease duration, Pdetmax, bladder com-pliance, and MCC as the core factors associated with UUTD risk. These findings add to the evidence on UUTD risk assessment, and the model provides a practical assessment tool.
Among the ML algorithms evaluated, KNN achieved the highest AUC in the training set, but its performance declined in the validation and test sets, suggesting potential overfitting. GNB, by comparison, performed consistently across all sets, with AUCs of 0.884, 0.868, and 0.878 in the training, validation, and test sets, respectively. Age, bladder management methods, MCC, and DSD differed significantly among the 3 sets. These results indicated stable GNB performance that was unaffected by the observed differences in patient characteristics under these conditions.
Bladder management method was an important risk factor for UUTD in our model. Abdominal pressure-assisted urination, in particular, was associated with a higher risk than other methods, such as CIC. This method involves applying manual pressure to the lower abdomen to aid bladder emptying, which can increase intravesical pressure during voiding. Chronic or recurrent elevations in intravesical pressure can damage the vesicoureteral junction, resulting in VUR and subsequent upper urinary tract damage. Other bladder management methods, in contrast, maintain lower, more stable intravesical pressures and thereby reduce UUTD risk.
DSD was also identified as a significant risk factor for UUTD in our model. In DSD, the urethral sphincter contracts involuntarily during detrusor contraction, causing functional outlet obstruction. The resulting increase in intravesical pressure during voiding can damage the upper urinary tract over time. Whereas abdominal pressure-assisted urination raises pressure externally, the elevation in intravesical pressure associated with DSD results from intrinsic neurological dysfunction. Early recognition of DSD and treatment with α-adrenergic receptor antagonists, intraurethral botulinum toxin injection, or other interventions can relieve DSD and prevent UUTD.
Disease duration was associated with an increased risk of UUTD in our model, consistent with previous studies [8, 26]. This association likely involves progressive remodeling and fibrosis of the bladder wall as chronic bladder dysfunction persists [27, 28]. Recurrent urinary tract infections, which become more frequent with longer NB duration, can further accelerate bladder wall fibrosis and damage the upper urinary tract [8]. These findings emphasize the importance of long-term followup and regular monitoring of patients with NB to detect UUTD early and allow prompt intervention.
Pdetmax, bladder compliance, and MCC each contributed to UUTD risk through specific pathophysiological mechanisms. In this study, Pdetmax was the value obtained during the filling phase. Elevated filling-phase Pdetmax indicates increased resting detrusor tone and can lead to persistently high intravesical pressure during bladder storage. Over time, this pressure elevation can impair ureteral peristalsis and cause hydronephrosis or renal insufficiency, thereby damaging the upper urinary tract [29]. Bladder compliance measures the bladder’s capacity to stretch and accommodate urine. With low compliance, small increases in bladder volume cause large increases in intravesical pressure, which can damage the upper urinary tract during storage [3, 7]. MCC measures the greatest volume the bladder can hold. Reduced MCC is associated with impaired bladder compliance and progressive detrusor fibrosis; these changes disrupt normal low-pressure storage, transmit higher pressure to the upper urinary tract, and consequently lead to UUTD [26]. These findings underscore the complexity of UUTD pathogenesis and the need to assess multiple aspects of bladder function comprehensively in patients with NB.
Subgroup analyses by upper urinary tract imaging outcome, NB etiology, and age demonstrated model performance across different patient groups. Earlier studies evaluated UUTD risk either in patients with a specific etiology [5, 7, 16, 17, 26] or in those with various neurological etiologies [3, 6, 8, 25], without subgroup analyses by NB etiology, imaging outcome, and age. After developing the UUTD risk assessment model for all patients with NB, we divided patients into spinal cord injury, spinal dysplasia, and other etiological subgroups; GNB outperformed the other models in each subgroup. Analyses by imaging outcome and age yielded similar results. These findings suggest that the model can assess UUTD risk across different groups of patients with NB. Because VUR, hydronephrosis, and reduced glomerular filtration rate are all clinically important manifestations of UUTD, we used a composite outcome to identify UUTD as comprehensively as possible; the relatively better model performance with the composite outcome than with a single modality indirectly reflected this more comprehensive assessment.
SHAP analysis was used to interpret model performance and assess the relative importance of each feature. By showing how individual features contribute to UUTD risk assessment, SHAP analysis may help clinicians develop personalized management plans for patients with NB. For example, patients who use ab-dominal pressure-assisted urination and have DSD require an appropriate bladder management method, such as CIC, to reduce intravesical pressure, whereas those with long disease duration and low bladder compliance can benefit from closer monitoring and timely intervention.
To our knowledge, this is the largest original study of UUTD risk assessment in patients with NB to confirm upper urinary tract outcomes using at least 2 imaging modalities and to incorporate both clinical and urodynamic variables. It is also the first to comprehensively evaluate UUTD risk models across different subgroups of patients with NB. The study nevertheless has several limitations. First, its retrospective, single-center design and restriction to patients under-going VUDS may introduce selection bias and limit generalizability. Although VUDS was indicated for routine evaluation and diagnosis of NB in this study, its use in clinical practice may vary with factors such as disease severity, economic status, and adherence. Because data were collected retrospectively from medical records, some relevant clinical parameters may not have been included. Second, although the study included a range of variables, other potential risk factors, such as radiomic features and medication use, were omitted for certain reasons. Future models should consider these factors to improve performance. Third, although patients came from multiple regions of our country, the geographically distinct test set provided internal-external validation and did not constitute independent external validation; independent multicenter or international validation is warranted. Fourth, most patients with NB underwent VUDS only once at our center, preventing assessment of changes in urinary tract function over time. Prospective longitudinal studies would help further validate the model’s capabilities and represent a key direction for future research.
In conclusion, we developed an ML model that integrates clinical and urodynamic parameters to assess UUTD risk in patients with NB. The findings emphasize the importance of bladder management method and urodynamic evaluation in this population, and the model provides a practical tool for UUTD risk assessment. Independent external validation of the model remains necessary.
Notes
Grant/Fund Support
This study was supported by the Key Technologies Research and Development Program (2023YFC3605300 and 2023YFC3605301) and the Fundamental Research Funds for Central Public Welfare Research (2023CZ-1, CRSI2025ZD-1-6).
Research Ethics
This study was approved by the institutional review board at the China Rehabilitation Research Center (No. 2023-075-01) and complied with the Declaration of Helsinki; this study was a retrospective study, and the data were processed to remove relevant private in-formation of patients, so the ethics committee agreed to exempt informed consent.
Conflict of Interest
No potential conflict of interest relevant to this article was reported.
AUTHOR CONTRIBUTION STATEMENT
· Conceptualization: LL
· Data curation: PZ, XL, TX
· Formal analysis: XL, TX
· Funding acquisition: LL
· Methodology: PZ, JW
· Project administration: LL, JW
· Visualization: XL, TX
· Writing - original draft: PZ
· Writing - review & editing: LL
