Machine Learning Model Predicts Survival in Rare Aggressive Bladder Cancer

Bladder neuroendocrine carcinoma (BNEC) is one of the rarest and most lethal malignancies urologists encounter, accounting for less than one percent of all bladder cancers yet behaving far more aggressively than the common urothelial carcinoma that dominates the field. Because the disease is so uncommon, treatment standards have been borrowed largely from small cell lung cancer, and prognostic assessment has relied almost entirely on the American Joint Committee on Cancer TNM staging system. That system, however, ignores patient demographics and treatment patterns, leaving clinicians with limited predictive power when counselling patients about their futures. A new retrospective study published in Frontiers in Oncology set out to close that gap by building and comparing eight machine learning models trained on nearly two decades of population-level data from the United States.

The research team, led by investigators at Zaozhuang Municipal Hospital and collaborating institutions in Shandong, China, extracted records of pathologically confirmed BNEC patients diagnosed between 2004 and 2021 from the Surveillance, Epidemiology, and End Results (SEER) database of the National Cancer Institute. From 2,858 candidate cases, 2,117 patients aged 18 or older with complete clinical information and at least one month of follow-up made it into the final cohort. The dataset captured demographic characteristics such as age, sex, and race; oncological features including primary tumour site, histological subtype, and AJCC and TNM stage; and treatment information covering surgery, radiotherapy, and chemotherapy. Patients were randomly divided in a seven-to-three ratio into a training set of 1,482 individuals and a validation set of 635, with no statistically significant differences between the two groups across any baseline variable.

The cohort profile underscored just how aggressive this disease is. The median age at diagnosis was 74 years, roughly four out of five patients were men, and 88.6 percent were white. Small cell carcinoma was by far the most common histological subtype, representing 76.9 percent of cases. Nearly half of all patients, 46.1 percent, already had stage III or IV disease at diagnosis, and 80.4 percent presented with tumours classified as T2 or beyond. Treatment patterns reflected both the rarity and the severity of the condition: 67.9 percent of patients underwent transurethral resection of the bladder tumour (TURBT), 24.7 percent received radical cystectomy, 61.3 percent received chemotherapy, and 23.5 percent received radiotherapy.

Before any machine learning could begin, the researchers used univariate and multivariate Cox regression to identify which factors independently shaped overall survival. Age above 74 years carried a hazard ratio of 1.51, meaning those patients faced roughly one and a half times the risk of death compared with younger patients. Male sex was associated with a modestly elevated risk, with a hazard ratio of 1.17. Tumour stage mattered enormously: patients with T4 tumours had a hazard ratio of 1.77 relative to those with T0, Ta, or T1 disease, while N2 nodal involvement carried a hazard ratio of 1.55. The strongest single predictor was distant metastasis, with M1 disease carrying a hazard ratio of 1.99. On the protective side, both surgical approaches were associated with better survival, TURBT with a hazard ratio of 0.75 and radical cystectomy with a hazard ratio of 0.36, as were radiotherapy at 0.79 and chemotherapy at 0.43. Carcinoid histology showed a strikingly low hazard ratio of 0.09, but the authors cautioned that the tiny number of such cases makes this an exploratory finding requiring cautious interpretation.

With those independent prognostic factors in hand, the team constructed eight machine learning survival models: gradient boosting machine (GBM), random survival forest, LASSO-Cox, CoxBoost, SurvivalSVM, XGBoost, supervised principal components, and partial least squares regression for Cox models. Each algorithm approaches the prediction problem differently. Gradient boosting builds an ensemble of weak learners sequentially, each correcting the errors of its predecessor, and was fitted here with a Cox proportional hazards objective. Random survival forests split patient groups using a log-rank splitting rule, while LASSO-Cox applies L1 regularization to shrink less informative variables toward zero, with the penalty parameter chosen by ten-fold cross-validation. XGBoost likewise used a Cox proportional hazards objective, and SurvivalSVM relied on a linear kernel with quadratic programming optimization.

Performance was evaluated using the concordance index and time-dependent receiver operating characteristic analysis, both corrected for right-censored observations through inverse probability of censoring weighting. In the training cohort, GBM led the field with time-dependent AUC values of 0.846 for one-year survival, 0.825 for three-year survival, and 0.816 for five-year survival. Random survival forest and plsRcox came close behind, while SurvivalSVM and SuperPC performed notably worse at every time point. In the validation cohort, GBM again topped the rankings with AUC values of 0.737, 0.730, and 0.742 at one, three, and five years respectively, ahead of LASSO-Cox, CoxBoost, plsRcox, and XGBoost. The GBM model also posted the highest C-index values in both cohorts, remaining above 0.69 at the sixty-month mark, and its performance curves tracked closely between training and validation sets, a sign of consistent rather than overfitted behaviour.

Crucially, the machine learning model outperformed the conventional TNM staging approach on the same validation cohort. Time-dependent ROC analysis showed lower AUC values for the TNM-based model at all three time horizons, suggesting that integrating demographics and treatment variables adds prognostic information that staging alone cannot capture. The authors noted that previously published BNEC nomograms achieved C-index values between 0.702 and 0.724, whereas the GBM model in this study demonstrated greater predictive accuracy, an advantage they attribute to the algorithm’s capacity to handle complex nonlinear relationships and interactions among variables.

Statistical discrimination alone does not guarantee clinical usefulness, so the team ran two further tests. Calibration curves showed that predicted survival probabilities clustered tightly around the diagonal reference line at one, three, and five years, with only a slight systematic underestimation in the low-probability range at the five-year horizon, a deviation the authors judged clinically acceptable. Decision curve analysis then asked the more practical question of whether acting on the model’s predictions would actually benefit patients. The GBM model’s net benefit curve exceeded both reference strategies across broad threshold ranges: 0.05 to 0.75 for one-year predictions, 0.10 to 0.70 for three-year predictions, and 0.15 to 0.65 for five-year predictions in the training set, with a modestly narrower but still clinically relevant range in the validation set. At a threshold probability of 0.3 for one-year survival, the net benefit reached approximately 0.35, peaking within the moderate-risk zone where clinical decisions matter most.

The model’s risk stratification ability provided perhaps the most intuitive demonstration of its value. Patients were split into high-risk and low-risk groups based on the median GBM-derived risk score, and Kaplan-Meier analysis revealed survival curves that separated from the very start of follow-up and maintained a substantial gap throughout, with a log-rank p value below 0.0001. Low-risk patients showed significantly prolonged median survival, confirming that the model can meaningfully sort patients into prognostically distinct subgroups rather than merely producing statistically impressive numbers.

The authors are careful about the limits of their work. The retrospective design and the nature of the SEER database introduce potential selection and information biases. The database lacks variables that could sharpen predictions further, including specific chemotherapy regimens, radiotherapy doses, surgical margin status, and molecular markers such as the Ki-67 proliferation index. No independent external validation was performed, and because BNEC is so rare, no comparably sized external cohort was available. The team also warns that associations between treatment variables and survival may be distorted by confounding by indication and immortal time bias, so they should not be read as causal treatment effects. Looking ahead, the researchers plan to validate the model in multicentre external cohorts and to integrate multimodal data such as radiomics and genomics, including TP53 and RB1 mutation status, in pursuit of genuinely precision-medicine-grade prognostication for this rare and unforgiving cancer.

Subject of Research: Machine learning-based survival prediction for bladder neuroendocrine carcinoma using SEER database data

Article Title: Development and validation of a machine learning-based model for predicting survival in patients with bladder neuroendocrine carcinoma: a large retrospective study

Article References: Qi, S., Dong, W., Shi, Z., Sun, Y., Zhang, D., Liu, C., Li, H., & Hu, L. (2026). Development and validation of a machine learning-based model for predicting survival in patients with bladder neuroendocrine carcinoma: a large retrospective study. Frontiers in Oncology, 16, 1956255. https://doi.org/10.3389/fonc.2026.1956255

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1956255

Keywords: bladder neuroendocrine carcinoma, machine learning, gradient boosting machine, SEER database, survival prediction, prognostic model, TNM staging, retrospective study, Cox regression, decision curve analysis, risk stratification, rare cancer