Bladder cancer ranks ninth in incidence and thirteenth in mortality among all malignancies. According to the Global Cancer Observatory, an estimated 613,791 new cases and 220,349 deaths were reported in 2022 (1). As the global population ages, the incidence of bladder cancer is projected to increase, posing continuous challenges to public health systems (2). In China, the improvement of medical record systems is exemplified by comprehensive records in Shanghai. The Shanghai Cancer Registry, renowned for its extensive scale and high-quality data, provides reliable information for analyzing cancer patient survival, mortality, and follow-up (3, 4). Tumor registry data from Putuo District offers a representative sample, facilitating effective in-depth analyses of bladder cancer.

Bladder cancer is recognized as one of the most economically demanding malignancies due to its complex treatment requirements and high recurrence rates (5, 6). Clinically, bladder cancer is classified into two distinct types: non-muscle-invasive bladder cancer (NMIBC) and muscle-invasive bladder cancer (MIBC), each with different management approaches and outcomes. The treatment typically involves substantial medical resources, including surgery, chemotherapy, and radiotherapy, resulting in significant direct medical costs. NMIBC management generally encompasses transurethral resection and intravesical therapy, whereas MIBC necessitates more aggressive interventions such as radical cystectomy combined with neoadjuvant or adjuvant systemic therapies. In addition to these costs, the disease imposes considerable socioeconomic impacts through lost productivity, prolonged unemployment, and the necessity for postoperative care. Furthermore, the management of bladder cancer patients involves periodic surveillance and ongoing medical care, which consume extensive healthcare and public health resources (7, 8). Surveillance is essential for ensuring treatment efficacy and monitoring disease progression, forming the basis for evaluating and adjusting therapeutic plans. However, these procedures significantly increase the complexity and cost of care. Thus, effective oncological management strategies must balance alleviating the financial burden on patients and their families with maintaining high-quality treatment outcomes. This necessitates healthcare providers and policymakers to optimize resource allocation, refine disease management strategies, and enhance the efficiency of healthcare services.

Accurate survival predictions enable the development of personalized follow-up plans, the optimization of examination schedules, and the organization of rehabilitation activities for bladder cancer patients. Individuals with a favorable prognosis might benefit from extended follow-up intervals to minimize unnecessary medical interventions, whereas high-risk patients necessitate intensive monitoring to detect recurrences or metastases promptly, thus ensuring timely treatment. Existing models for predicting bladder cancer survival rates exhibit varying predictive values and influencing factors. Traditional linear models, such as logistic regression, encounter limitations when handling large datasets and often fail to capture complex variable interactions (9, 10). Machine learning techniques have emerged to overcome these challenges, employing advanced algorithms to analyze multidimensional data patterns, including nonlinear relationships, thereby enhancing predictive accuracy (11).

Deep learning, as a subset of machine learning, has substantially advanced predictive modeling. These models excel in processing data and offer remarkable flexibility by automatically extracting and utilizing complex features, thereby enhancing predictive accuracy and generalization (12). Among these, the TabNet model stands out as a novel approach for tabular data, combining the strengths of decision trees and neural networks. This integration provides robust interpretability and efficient feature utilization, making it a powerful tool in predictive modeling (13).

The TabNet model has been successfully applied to prognostic analysis of liver metastases in rectal cancer patients, demonstrating high accuracy and reliability (14). Its application in bladder cancer survival analysis, however, has not yet been explored. This study employs the TabNet model to predict 5-year overall survival (OS) and cancer-specific survival (CSS) for bladder cancer patients in Putuo District, and to identify key risk factors affecting survival rates. This research is expected to provide a more precise prognostic tool for bladder cancer, thereby supporting public health decisions, optimizing follow-up schedules, and informing treatment strategies.

Bladder cancer, the second most common malignancy in the genitourinary system after prostate cancer, has rising incidence and mortality rates, imposing significant burdens on society and healthcare systems (17). Effective predictive models are crucial for community-based cancer management, patient follow-up, and healthcare resource allocation. This study utilized data from the Putuo District within the Shanghai Cancer Registration and Reporting System to develop and validate an individualized survival prediction model for bladder cancer patients using deep learning algorithms. The TabNet model demonstrated enhanced performance in predicting 5-year OS and CSS compared to logistic regression, achieving higher accuracy and better calibration. In the validation set, TabNet demonstrated superior performance in ROC curves, decision curve analysis, and calibration curves, indicating high clinical utility. SHAP analysis identified age, N stage, and T stage as the primary factors affecting 5-year OS and CSS. These findings provide a basis for improving community-based cancer surveillance and facilitating more efficient population health management for bladder cancer patients.

The 5-year survival rates for bladder cancer patients exhibit regional variations globally. Europe reports an age-standardized relative 5-year OS of approximately 70% (18). In the United States, the average 5-year OS is 77%, with 19% of patients succumbing to bladder cancer (19). Data from China show an increase in the 5-year survival rate from 67.3% to 71.6% between 2003 and 2015, comparable to international levels (20). This study found a 5-year OS rate of 70.48% and a CSS rate of 81.94%, slightly lower but generally in line with the national data. According to the National Central Cancer Registry (NCCR) of China (21), 91.4% of bladder cancers are urothelial carcinoma, with approximately 55-60% classified as low-grade tumors. These proportions align with our findings, indicating the representativeness and reliability of the sample. The cancer registry provides high-quality, comprehensive data, enabling effective monitoring of cancer incidence and survival rates, thereby supporting the development and evaluation of cancer prevention and control strategies.

Traditionally, clinicians have relied on the American Joint Committee on Cancer (AJCC) staging guidelines, using T, N, and M stages for preliminary prognosis assessment (22). However, this system fails to account for demographic factors or treatment modalities, which significantly influence outcome prediction. Consequently, researchers have developed more comprehensive predictive models. Zhang et al. (23). integrated TNM staging with clinical parameters to construct a nomogram for visualizing survival rates, achieving a C-index of 0.813 for 5-year OS, thereby demonstrating superior predictive performance compared to the AJCC-TNM staging. Similarly, He et al. (24). developed a nomogram based on the Surveillance, Epidemiology, and End Results (SEER) database to predict CSS for postoperative bladder cancer at the population level, yielding a C-index of 0.823. Despite the acceptable predictive accuracy of these models, certain limitations exist. The complexity of the calculations and the challenges in addressing collinearity among variables may impact the stability of the results. In this study, logistic regression analysis was used to predict 5-year OS, resulting in an AUC of 0.768, and 5-year CSS, resulting in an AUC of 0.742. These findings are consistent with previous research, indicating a moderate level of predictive capability and highlighting the potential for further improvement.

The continuous advancement of artificial intelligence has led to the development of innovative predictive tools. Leveraging multi-omics data from large-scale databases (GEO, TCGA, and IMvigor210), ensemble machine learning approaches have successfully constructed robust risk stratification models, demonstrating significant value in predicting treatment response and immunotherapy outcomes for bladder cancer patients (25). In a study involving 161,227 bladder cancer patients, Bhambhvani et al. (26). employed clinical-pathological data and sociodemographic variables to train an artificial neural network (ANN) for predicting 5-year CSS, achieving an AUC of 0.81, which was more accurate than traditional multivariate models. Although ANNs are proficient at capturing complex patterns and nonlinear relationships, they present challenges in transparency, feature selection, and susceptibility to overfitting. In comparison, deep learning utilizes more intricate neural network structures to enhance the extraction and processing of complex features. These advanced models excel in handling high-dimensional and large-scale data, effectively mitigating overfitting risks (27, 28). In this study, the TabNet deep learning model demonstrated AUCs of 0.874 for 5-year OS and 0.864 for CSS, significantly exceeding the performance of logistic regression models. The moderate F1 and Kappa values reflect the sensitivity of these metrics to class imbalance (80% survival vs. 20% death), while the high AUC and positive decision curve analysis confirm strong discrimination ability for risk stratification. These findings underscore the potential of TabNet in prognostic prediction for bladder cancer and provide strong support for its application in community-based cancer management and surveillance. Indeed, the attention mechanism in TabNet facilitates automatic feature selection, addresses collinearity issues, and enhances model transparency and interpretability, making it particularly suitable for population health management.

Bladder cancer is a complex disease influenced by various factors, significantly impacting survival rates. Age at diagnosis is a well-known predictor of poor OS and CSS in bladder cancer (29, 30). A study indicates that cancer mortality is 15 times higher in individuals aged ≥65 compared to those under 65 (31). In our study, age was identified as the most significant factor affecting 5-year OS. This can be attributed to the decline in immune function in elderly patients, reducing their resistance to tumors. The presence of chronic diseases further increases the risk of treatment complications. Additionally, older patients are often diagnosed at more advanced stages, delaying optimal treatment (32). In another study based on data from the SEER database, Wang et al. (33). found that age, T stage, and N stage were the important prognostic factors, which is consistent with our findings. The T stage indicates the extent of tumor invasion, with higher stages correlating with deeper bladder wall invasion and poorer prognosis (34). The N stage reflects lymph node involvement, with higher stages suggesting greater metastatic risk, complicating treatment, and lowering survival rates (35).

Moreover, our study identified smoking and tumor grade as additional prognostic factors with a considerable degree of influence. Numerous studies have demonstrated a strong association between smoking and the development of bladder cancer, identifying it as a significant risk factor (36–38). Approximately 30-50% of bladder cancer deaths are attributed to smoking across various populations (39). Although smoking status ranked as a moderate predictor in our SHAP analysis, its modifiable nature distinguishes it from other prognostic factors. This suggests that smoking cessation interventions integrated into survivorship care could potentially improve outcomes, representing a practical target for both clinical and community-level prevention strategies. Similar to 5-year OS, 5-year CSS in our study was primarily influenced by age, T stage, and N stage, with smoking and histological features also impacting patient prognosis. Patients with non-urothelial carcinomas, such as adenocarcinoma and squamous cell carcinoma, have poorer prognoses, likely due to the aggressive nature of these histological types and their poorer response to standard treatments (40).

Prognostic modeling in bladder cancer has achieved significant advances, with Li et al. (41) developing a validated nomogram for predicting overall survival in postoperative high-grade bladder urothelial carcinoma patients using SEER data and external validation cohorts, while Bhambhvani et al. demonstrated that ANN could achieve enhanced discriminative performance in 5-year survival prediction (26). In the present study, we employed TabNet architecture integrated with SHAP analysis to enhance model interpretability. The SHAP framework enables decomposition of model predictions into additive feature attributions, revealing not only which variables influence prognosis but also the magnitude and directionality of their effects across different patient profiles. This transparency facilitates personalized risk assessment by allowing clinicians to understand the specific factors driving individual patient predictions, thereby supporting evidence-based treatment planning and informed patient counseling. An ideal prognostic tool should incorporate readily available variables with significant predictive value while maintaining interpretability, and our TabNet model with SHAP analysis achieved satisfactory predictive performance alongside transparent feature attribution.

This study has several limitations. First, the lack of data on surgical methods, patient comorbidities, and socioeconomic status may affect prognostic assessment. Efforts were made to mitigate this limitation by incorporating a comprehensive set of other relevant variables to enhance the robustness of the analysis. Second, excluding patients with incomplete data may introduce selection bias. However, this approach was necessary to maintain data integrity and analytical accuracy, thereby preventing potential errors due to missing data. Third, this model was developed and validated exclusively using data from Putuo District. Although the Shanghai Cancer Registry ensures high-quality population-based surveillance with comprehensive follow-up, the geographic confinement to a single district raises concerns about generalizability. Regional variations in bladder cancer epidemiology, including risk factor distributions and clinical presentations, may influence model performance across different populations. External validation using datasets from geographically and demographically diverse populations is essential to assess the model’s transferability and clinical utility. Prospective validation efforts would help determine the model’s broader applicability and inform necessary adaptations for specific regional contexts. Fourth, our cohort of 620 patients is relatively small for deep learning. However, rigorous validation demonstrated satisfactory generalization, and TabNet’s superior performance over logistic regression justified this approach. Larger multi-center datasets would further strengthen model robustness and enable broader generalizability assessment. Fifth, our comparison was limited to TabNet and logistic regression. Comprehensive benchmarking with additional machine learning methods (XGBoost, Random Forest, multilayer perceptron) would provide valuable insights and represents an important direction for future research with larger multi-center datasets.

The deep learning-based TabNet model shows great potential in predicting survival outcomes for bladder cancer patients, offering advantages such as handling high-dimensional data and providing robust performance. This model can be a valuable tool for community healthcare workers to identify high-risk patients who require more intensive follow-up and resource allocation. Age, T stage, and N stage emerged as the most significant prognostic factors, highlighting their critical role in determining patient outcomes and providing guidance for developing targeted community intervention strategies. External validation across diverse populations is warranted to establish the model’s generalizability before broader clinical implementation.