We collected inpatient data from 1,764 patients diagnosed with bladder cancer and undergoing radical cystectomy at the First Affiliated Hospital and the Second Affiliated Hospital of Kunming Medical University between 2014 and 2024. The dataset included demographic characteristics (gender, age, ethnicity, weight, smoking, alcohol consumption, etc.), medical history (hypertension, diabetes, hematuria, frequency, urgency, dysuria, difficulty in urination, and previous surgeries), tumor morphological features (tumor shape, diameter, location, number, presence of a base, boundary clarity, color, texture, presence of necrosis, bleeding, and cystic lesions), as well as pathological features assessed by immunohistochemistry, including UPK3A, GATA Binding Protein 3 (GATA3), Cytokeratin 20 (CK20), Cytokeratin 7 (CK7), Cytokeratin 5/6 (CK5/6), Tumor Protein 63 (P63), Tumor Protein 53 (P53), Androgen Receptor (AR), Programmed Death-Ligand 1 (PD-L1), microsatellite stability, Human Epidermal Growth Factor Receptor 2 (HER2), nerve invasion, vascular invasion, pathological staging, grading, positive surgical margins, and histological types such as squamous, glandular, neuroendocrine, and sarcomatoid.

Inclusion criteria for participants were as follows: (1) patients aged over 18 years; (2) diagnosis of bladder cancer according to the WHO Classification of Tumors of the Urinary and Male Genital Systems (4th Edition), and receipt of radical cystectomy; (3) complete clinical data, including blood count, biochemical tests, pathological parameters, and immunohistochemistry; (4) detailed treatment history with complete follow-up data and results; (5) no prior radiotherapy, chemotherapy, or immunotherapy.

Patients who met any of the following criteria were excluded from the study: (1) patients who underwent partial bladder tumor resection; (2) post-operative pathology confirmed non-urothelial carcinoma; (3) incomplete clinical data or lost to follow-up with no available prognostic data; (4) preoperative radiotherapy, chemotherapy, or immunotherapy; (5) other malignancies metastasized to the bladder; (6) patients under 18 years of age; (7) patients with survival time less than 1 month. The study was approved by the Ethics Committees of the First Affiliated Hospital and the Second Affiliated Hospital of Kunming Medical University, with informed consent obtained from all patients.

To minimize the impact of missing data on model construction, we used the K-Nearest Neighbors (KNN) Imputer method to impute missing data (less than 20% missing), while data with more than 20% missing were excluded. The primary endpoint was the response to postoperative adjuvant therapy, as recorded in the patients’ follow-up treatment records. Missing values were imputed using the `KNNImputer` algorithm (version 0.24.2, scikit-learn), with the number of neighbors set to 5. Variables with more than 20% missing data were excluded from model construction to reduce bias. Continuous variables were standardized using z-score normalization, and categorical variables were encoded using one-hot encoding prior to model input.

Categorical variables are presented as percentages (%) and compared between groups using Pearson’s chi-square test. Due to the imbalance in the dependent variable categories, an undersampling method was applied to resample the data and balance the distribution. A five-fold cross-validation was used to split the dataset into training and internal validation sets. In the case of high-dimensional features, LASSO regression was employed for feature selection. This method applies L1 regularization to shrink regression coefficients, reducing dimensionality, selecting the most informative variables, and eliminating redundant features.

Nine ML algorithms were used for predictive modeling, including XGBoost, support vector machine (SVM), multilayer perceptron (MLP), KNN, logistic regression, LASSO, decision tree (DT), gradient boosting machine (GBM), and RF. All models incorporated the features selected by LASSO. A single cross-validation was performed to ensure model stability. Grid search optimization was applied to fine-tune hyperparameters, and the model with the highest area under the AUC-ROC curve was selected as the optimal model. The final model was constructed on the training set and validated on both internal and external validation sets. Model performance was evaluated using AUC-ROC, sensitivity, specificity, recall, F1 score, and accuracy. All machine learning models were implemented using Python 3.8 with `scikit-learn` (v0.24.2), `xgboost` (v1.5.0), `lightgbm` (v3.3.1), and `shap` (v0.41.0). LASSO regression for feature selection was performed using `LassoCV` from `scikit-learn`, with 10-fold cross-validation to determine the optimal lambda value (λ = 0.009), minimizing binomial deviance. Model hyperparameters (e.g., learning_rate, n_estimators, max_depth) were tuned using `GridSearchCV` with five-fold cross-validation. The hyperparameter configurations for each model are provided in Supplementary Table S1 .

Additionally, to assess the real clinical utility of the model, DCA and calibration curves were plotted. To identify the optimal clinical decision threshold, a clinical impact curve (CIC) was constructed to visually assess the most effective decision threshold. The threshold was derived using the “surv_cutpoint” function in the survminer R package to maximize survival difference. To analyze the impact of the selected features on the model predictions, SHAP analysis was used. SHAP summary plots were generated to show the contribution of each feature to the prediction results, and specific cases were evaluated using SHAP to illustrate the degree of impact of selected features on the predictions. All statistical analyses were conducted in Python, with two-sided p-values < 0.05 considered statistically significant.

(The flowchart of this research is shown in Figure 1 )

All computational procedures were implemented using Python 3.8. Machine learning models, including LightGBM, XGBoost, Random Forest, SVM, and others, were built using the scikit-learn (v0.24.2), xgboost (v1.5.0), and lightgbm (v3.3.1) packages. Data imputation was performed using KNNImputer with default parameters (n_neighbors=5), and variables with more than 20% missing data were excluded. Continuous variables were standardized using z-score normalization, and categorical variables were transformed by one-hot encoding.

LASSO regression for feature selection was conducted using LassoCV from scikit-learn with 10-fold cross-validation to determine the optimal penalty parameter (λ = 0.009), based on minimum binomial deviance. Model hyperparameters were optimized via GridSearchCV with five-fold internal cross-validation. The detailed hyperparameter settings for each model are provided in Supplementary Table S1 .

Model performance was evaluated using metrics such as accuracy, sensitivity, specificity, precision, negative predictive value (NPV), F1-score, Youden index, and AUC-ROC. All metrics were computed using functions from scikit-learn.metrics. To interpret feature contributions, SHAP (SHapley Additive exPlanations) values were calculated using the TreeExplainer module from the shap Python package (v0.41.0). Both global summary plots and individual force plots were generated to visualize model decision logic and feature importance.

All analytical pipelines, including data preprocessing, model training, evaluation, and SHAP analysis, were version-controlled and archived. The complete source code and training-validation splits are available from the corresponding author upon reasonable request, ensuring full reproducibility.

The complete analytical process of this study is illustrated in Figure 1 . In this study, based on data from 1,674 bladder cancer patients at two affiliated hospitals of Kunming Medical University, a final cohort of 1,032 eligible cases was included. These cases were randomly divided into a training cohort (N=412) and a validation cohort (N=620) in a 4:6 ratio. Univariate analysis identified clinical features associated with patient outcomes, including age, urinary urgency, dysuria, tumor necrosis, perineural invasion, vascular invasion, tumor diameter, tumor location, tissue texture, as well as several blood and urine markers (such as creatinine, neutrophil count, and leukocyte esterase), which showed significant differences between outcome groups (P<0.05) ( Table 1 ).

Table 1 : Baseline UPK3A, Serum and Demographic Date in Bladder cancer (Please refer to the attached file ( Table 1 . DOCX) for details).

Subsequently, clinical variables with statistical significance were selected through univariate analysis, and LASSO regression was applied for further feature selection. The LASSO path plot ( Figure 2A ) shows that as the regularization parameter λ increases, the regression coefficients of some features gradually converge to zero. The cross-validation curve ( Figure 2B ) determined that the optimal λ value was 0.009, which minimized the binomial deviance of the model. The final selected features included age, smoking history, positive urine bacterial culture, perineural invasion, vascular invasion, muscle layer invasion (M stage), UPK3A expression, tumor number, tumor boundary characteristics, and necrosis ( Figure 2C ). These features provided an essential foundation for subsequent model construction.

Among these factors, smoking history, urine leukocytes, and vascular invasion were considered key prognostic factors for BUC, as these features are likely closely related to the mechanisms of cancer development and progression. For example, smoking, a known risk factor for bladder cancer, may contribute to the malignant transformation of urothelial cells through the accumulation of carcinogens, while an increase in urine leukocytes may suggest the role of inflammatory responses in tumor progression.

This study constructed multiple ML models (KNN, RF, XGBoost (XGB),SVM, Logistic Regression (LR), MLP, LightGBM, LASSO, and DT to predict the prognosis of BUC. The models’ performance was systematically evaluated using ROC curves, Calibration Curves, and DCA to identify the optimal predictive model.

Our results indicate that in both the training and validation sets, LightGBM, RF, and XGB models demonstrated excellent predictive performance. The training set AUCs were 0.894, 0.894, and 0.872, respectively ( Figure 3A , Table 2 ), and the validation set AUCs were 0.741, 0.754, and 0.751, respectively ( Figure 3B , Table 2 ). LightGBM and RF also outperformed other models in terms of Accuracy, Recall, and F1-Score. The confusion matrix ( Table 3 ) further validated the stability of the models in classifying true positives (TP) and true negatives (TN).

The calibration performance of the models was assessed through Calibration Curves, and most models, particularly LightGBM and RF, showed good agreement between predicted probabilities and actual observations in both the training and validation sets ( Figures 3C, D ). DCA ( Figures 3E, F ) further demonstrated that LightGBM, RF, and XGB models provided higher net clinical benefits at various probability thresholds, suggesting that these models have substantial potential for practical clinical application.

In summary, this study constructed and validated a series of machine learning-based prognostic prediction models for BUC. XGBoost and LightGBM exhibited superior performance in classification (AUC), calibration (accuracy of predicted probabilities), and clinical DCA, making them suitable for prognostic prediction and personalized risk assessment in BUC patients. These results provide clinicians with effective risk stratification tools, helping to more accurately identify high-risk patients and formulate individualized treatment strategies.

The clinical application value of different ML models in predicting the prognosis of BUC was further evaluated using CIC. CIC primarily illustrates the number of patients predicted as high-risk at various risk thresholds and the number of those who actually experience the target event (e.g., disease recurrence or progression).

The analysis in this study shows that the LightGBM,RF, and XGBoost models were able to accurately identify a higher number of high-risk patients across various risk thresholds. Furthermore, the proportion of actual events (e.g., disease recurrence or progression) occurring among those predicted as high-risk was higher, with the curve trends closely mirroring the actual event occurrence curve, indicating their higher clinical application value. In contrast, the KNN and DT models showed considerable deviation from the actual results at medium and low-risk thresholds, with lower accuracy.

Overall, both RF and LightGBM models maintained a good balance between sensitivity and specificity across all risk thresholds, demonstrating the optimal clinical net benefit. These findings support the use of RF and LightGBM as the preferred models for prognostic risk stratification in BUC patients ( Figure 4 ).

To enhance the interpretability of the model, SHAP analysis was applied in this study. This method quantifies the contribution of each feature to the model’s decision-making process, further exploring its clinical significance. SHAP values reflect the direction and magnitude of each variable’s impact on the model’s output, where positive SHAP values indicate that the variable increases the probability of predicting a high-risk outcome, and negative SHAP values reduce this probability.

The analysis revealed that, in the RF model, features such as vascular invasion, perineural invasion, muscle layer infiltration (M stage), tumor necrosis, tumor boundary clarity, urine leukocyte esterase positivity, and white blood cell count had the greatest impact on the model’s predictions. The direction of change in feature values was strongly correlated with the predicted risk levels ( Figure 5 ). These findings not only enhance the biological interpretability of the model but also provide a theoretical basis for the postoperative management of bladder cancer patients in the future.