Machine Learning Model Predicts Survival in Rare Aggressive Bladder Cancer
Bladder neuroendocrine carcinoma (BNEC) is one of the rarest and most deadly cancers that doctors who treat bladder diseases encounter. It makes up less than one percent of all bladder cancers, but it grows and spreads much more aggressively than the common type, called urothelial carcinoma. Because BNEC is so rare, doctors have mostly based its treatment on methods used for small cell lung cancer. They have also relied almost entirely on the American Joint Committee on Cancer TNM staging system to estimate how long patients might survive. However, this system does not consider factors such as patients' personal characteristics or the treatments they receive. This limits doctors' ability to predict what may happen to individual patients. A new study published in Frontiers in Oncology aimed to address this problem by building and comparing eight machine learning models using nearly 20 years of population data from the United States.
The research team included scientists from Zaozhuang Municipal Hospital and other institutions in Shandong, China. They collected records from the National Cancer Institute's Surveillance, Epidemiology, and End Results (SEER) database for patients whose BNEC diagnosis had been confirmed by examining tissue samples between 2004 and 2021. Of 2,858 possible cases, 2,117 patients aged 18 or older were included in the final study. These patients had complete medical information and had been followed for at least one month. The data included personal details such as age, sex, and race; cancer details such as the tumour's location, cell type, and AJCC and TNM stages; and treatments such as surgery, radiotherapy, and chemotherapy. The researchers randomly divided the patients into two groups in a ratio of seven to three. The training group contained 1,482 patients, and the validation group contained 635. The groups showed no statistically significant differences in any of the starting characteristics measured by the researchers.
The patient data showed just how aggressive this cancer can be. The median age at diagnosis was 74, meaning half the patients were older than 74 and half were younger. About four out of every five patients were men, and 88.6 percent were white. Small cell carcinoma was by far the most common cancer type, accounting for 76.9 percent of cases. Nearly half of the patients, 46.1 percent, already had stage III or IV cancer when diagnosed. Also, 80.4 percent had tumours classified as T2 or higher, meaning the tumours had grown into deeper layers of tissue. Treatment patterns reflected the rarity and seriousness of the disease: 67.9 percent of patients underwent transurethral resection of the bladder tumour (TURBT), a procedure that removes a tumour through the urethra; 24.7 percent had their bladder removed through an operation called radical cystectomy; 61.3 percent received chemotherapy; and 23.5 percent received radiotherapy.
Before building the machine learning models, the researchers used two types of Cox regression analysis, called univariate and multivariate analysis, to find which factors were independently linked to overall survival. Being older than 74 was linked to a hazard ratio of 1.51, meaning these patients had about 1.5 times the risk of death compared with younger patients. Being male was linked to a slightly higher risk, with a hazard ratio of 1.17. The stage of the tumour also mattered greatly. Patients with T4 tumours had a hazard ratio of 1.77 compared with patients whose tumours were classified as T0, Ta, or T1. Patients with cancer in N2 lymph nodes had a hazard ratio of 1.55. The strongest single predictor was distant spread of cancer: patients with M1 disease had a hazard ratio of 1.99. Some treatments were linked to better survival. TURBT had a hazard ratio of 0.75, radical cystectomy had a hazard ratio of 0.36, radiotherapy had a hazard ratio of 0.79, and chemotherapy had a hazard ratio of 0.43. Carcinoid cancer cells were linked to a much lower hazard ratio of 0.09. However, the authors warned that very few patients had this cell type, so this finding is only an early clue and must be interpreted carefully.
Using the factors that were independently linked to survival, the team built eight machine learning 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 method works differently. Gradient boosting combines many simple prediction models in a series, with each new model trying to correct mistakes made by the previous ones. In this study, it used a Cox proportional hazards objective. Random survival forests divide patients into groups using a method based on the log-rank test. LASSO-Cox uses a mathematical penalty to reduce the influence of less useful variables, sometimes shrinking their effects to zero. The researchers chose the penalty setting through ten-fold cross-validation, which repeatedly tests the model on different parts of the data. XGBoost also used a Cox proportional hazards objective, while SurvivalSVM used a linear kernel and an optimization method called quadratic programming.
The researchers measured how well the models predicted survival using the concordance index and time-dependent receiver operating characteristic analysis. Both methods accounted for patients whose follow-up ended before their final survival outcome was known, using a technique called inverse probability of censoring weighting. In the training group, GBM performed best, with area under the curve (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 had results close behind, while SurvivalSVM and SuperPC performed noticeably worse at every time point. GBM also ranked first in the validation group, with AUC values of 0.737, 0.730, and 0.742 for one-, three-, and five-year survival, respectively. It performed better than LASSO-Cox, CoxBoost, plsRcox, and XGBoost. GBM also had the highest concordance index (C-index) in both groups, staying above 0.69 at the 60-month point. Its performance curves were similar in the training and validation groups, suggesting that its predictions were consistent rather than overly fitted to the training data.
Importantly, the machine learning model performed better than the traditional TNM staging method in the same validation group. Time-dependent ROC analysis showed that the TNM-based model had lower AUC values at all three time points. This suggests that adding information about patients' personal characteristics and treatments can help predict survival beyond what cancer stage alone can tell doctors. The authors also noted that earlier BNEC prediction charts, known as nomograms, had C-index values between 0.702 and 0.724. The GBM model in this study showed better predictive performance. The researchers believe this advantage may come from the algorithm's ability to recognize complex patterns, including relationships between variables that do not follow a simple straight-line pattern.
Good statistical performance does not automatically mean a model will help patients in real medical settings, so the researchers carried out two more tests. Calibration curves showed that the predicted chances of survival were close to the actual outcomes at one, three, and five years. At five years, the model slightly underestimated survival chances among patients with low predicted probabilities. The authors considered this difference small enough to be acceptable in clinical practice. The team also used decision curve analysis to examine whether using the model to guide decisions could benefit patients. The GBM model showed greater net benefit than both comparison strategies across wide ranges of decision thresholds. In the training group, these ranges were 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 validation group, the useful ranges were somewhat narrower but still clinically meaningful. For one-year survival at a threshold probability of 0.3, the net benefit was about 0.35. The highest benefit occurred in the moderate-risk range, where clinical decisions may be especially important.
The model's ability to separate patients into risk groups offered another clear sign of its potential value. The researchers divided patients into high-risk and low-risk groups using the middle value, or median, of the risk scores calculated by GBM. They then used Kaplan-Meier analysis to compare the groups' survival over time. The survival curves separated from the beginning of follow-up and remained far apart. The log-rank test produced a p value below 0.0001, showing strong evidence of a difference between the groups. Patients in the low-risk group had a significantly longer median survival time. This suggests that the model can meaningfully sort patients into groups with different expected outcomes, rather than simply producing impressive statistical results.
The authors clearly explained the limits of their research. Because the study looked back at existing records rather than following a group of patients in a planned experiment, and because of the way the SEER database collects information, the findings may be affected by selection bias and information bias. The database also lacks details that could improve predictions, including the exact chemotherapy regimens used, radiotherapy doses, whether surgery left any cancer cells behind at the edges of the removed tissue, and molecular markers such as the Ki-67 proliferation index, which measures how quickly cells are growing. The model was not tested in a separate external patient group, and BNEC is so rare that the researchers could not find another group of similar size for this purpose. The team also warned that the links between treatments and survival may be affected by confounding by indication and immortal time bias. These problems can make treatments appear more or less helpful than they really are, so the findings should not be taken as proof that a particular treatment directly causes better survival. In future work, the researchers plan to test the model in patient groups from multiple medical centres and combine different kinds of data, including radiomics and genomic information such as TP53 and RB1 mutation status. Their goal is to develop more precise survival predictions for this rare and highly aggressive cancer.
Subject of Research: Using machine learning and data from the SEER database to predict survival in patients with bladder neuroendocrine carcinoma
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