Predicting bladder cancer outcomes after surgery. This study applied machine learning and information theory to predict long-term cancer recurrence and survival in 3,503 bladder cancer patients who underwent radical cystectomy at the University of Southern California between 1971 and 2016.
Radical cystectomy -- surgical removal of the bladder, regional lymph nodes, and adjacent organs -- is the standard treatment for invasive bladder cancer. Accurately predicting which patients will experience cancer recurrence or reduced survival after this major surgery is essential for guiding post-operative treatment decisions.
The study used preoperative and operative clinical measurements, including imaging results, surgical findings, and pathology data, to train models capable of predicting 1-year, 3-year, and 5-year outcomes following surgery.
Key achievement: The machine learning models achieved greater than 70% sensitivity and specificity for predicting both recurrence and survival, outperforming traditional single-variable approaches and existing nomograms used in clinical practice.
TNM staging governs bladder cancer classification. Bladder cancer is staged using the TNM system, where the T stage reflects how deeply the tumor invades the bladder wall -- from Ta (noninvasive papillary) through T4 (invasion into adjacent organs).
Radical cystectomy outcomes are strongly influenced by surgical precision. Post-cystectomy survival is higher when negative surgical margins are achieved and more than ten pelvic lymph nodes are removed during the procedure. Patients with organ-confined disease rarely relapse beyond 5 years even without adjuvant treatment.
Current prognostic tools in clinical practice either reflect cohort-level outcomes rather than individual predictions, or are limited to a single 5-year recurrence estimate using nomograms. Machine learning offers an opportunity to provide higher-resolution, individualized risk stratification.
Prior machine learning efforts in cancer prediction have demonstrated promise, but penetration into routine clinical practice remains low. This study aimed to benchmark the predictive value of existing clinical data collection standards using a large, continuously maintained dataset.
A 45-year prospective dataset formed the foundation. The dataset included 3,499 patients (after removing 4 with missing survival data) who underwent radical cystectomy at USC from 1971 to 2016. Data covered demographics, preoperative clinical diagnostics, tumor markers, pathologic findings, surgical data, and adjuvant therapy information.
Mutual information (MI) -- a measure from information theory -- was used to assess pairwise correlations between all variables. This produced a network-based view of the entire clinical database, enabling identification of which predictors were most closely associated with long-term outcomes.
Feature selection reduced the full set of 73 predictors to between 42 and 54 variables per model by removing redundant and irrelevant predictors. Redundancy was addressed by hierarchical clustering, and irrelevance was addressed through MI thresholds against the target outcomes.
Six base machine learning classifiers were tested: support vector machines (SVM), bagged SVM, K-nearest neighbor (KNN), adaptive boosted trees (AdaBoost), random forest (RF), and gradient boosted trees (GBT). Ensemble meta-classifiers combining the best performing base models were then built using hard-voting and stacking approaches.
Class imbalance between recurring and non-recurring patients was addressed through random oversampling. Final performance was evaluated using nested 10-fold cross-validation with 5 inner folds for hyperparameter tuning, implemented in the Scikit-Learn Python library.
Disease status at surgery strongly predicts survival. Patients who never experienced recurrence had a mean overall survival of 7.93 years, while those who were never disease-free had a mean survival of just 0.81 years. Survival declined sharply in the first five years post-cystectomy before stabilizing.
Pathologic staging at the time of cystectomy provided much greater prognostic resolution than clinical staging performed before surgery. For example, pathologic staging clearly differentiated P3a from P2b patients, while clinical staging failed to separate T2b from T3a patients.
Clinical staging disagreed with pathologic staging in the majority of cases: during the full study period, clinical staging underestimated pathologic stage in 51% of patients and over-estimated it in 25%. This discrepancy has significant implications for pre-surgical treatment planning.
Disease extent at surgery defined survival tiers: 5-year survival rates were 75% for organ-confined disease, 41.3% for extra-vesical disease, and 21.3% for node-positive disease. Lymphovascular invasion further stratified outcomes, with 5-year survival of 29.9% versus 63.7% in patients with and without invasion, respectively.
Pathologic stage was the strongest predictor of both outcomes. Pathologic stage subgroup -- categorizing patients as organ-confined, extra-vesical, or node-positive -- had the highest chi-squared association with both 5-year recurrence and overall survival, followed closely by pT stage under TNM 5th and 7th editions.
Lymph node-related variables were particularly important for recurrence prediction. The number of positive lymph nodes removed ranked 7th overall and was more strongly correlated with recurrence than the total number of lymph nodes removed (ranked 27th). Pathologic lymphovascular invasion ranked 9th.
Age at cystectomy and type of urinary diversion were more strongly associated with overall survival than with recurrence, while neoadjuvant chemotherapy variables were more predictive of recurrence. This divergence reflects the biologically distinct drivers of cancer relapse versus patient mortality.
Comorbidity variables (such as diabetes) showed very low mutual information with cancer-specific variables and long-term outcomes, suggesting that pre-existing non-cancer conditions have limited direct impact on bladder cancer progression trajectories in this dataset.
Even lower-ranked predictors added value. Because pairwise mutual information among predictors was generally low, even variables with weaker individual correlations with outcomes contributed independent information about each patient, justifying the multivariate approach.
Meta-classifiers outperformed single-predictor models across all timepoints. For recurrence prediction, all meta-classifiers achieved sensitivities and specificities above 70%, while single-predictor logistic regression models showed imbalance between these metrics despite sometimes achieving higher sensitivity alone.
For recurrence, F1 scores improved from 0.508 at 1 year to 0.636 at 5 years, likely because the ratio of recurrence-positive to recurrence-negative patients becomes more balanced at longer follow-up periods, making classification easier.
Survival meta-classifiers showed even stronger performance than recurrence models. The 5-year survival meta-classifier achieved sensitivity of 0.741, specificity of 0.768, precision of 0.780, and an F1 score of 0.760 -- substantially better than single-predictor models for the same timepoint.
The 1-year meta-classifiers provided the best combination of sensitivity and specificity for both outcomes, while later-year models offered higher precision, making them useful for distinguishing lower- and higher-risk patients within the first 5 years after surgery.
Machine learning can complement and exceed current prognostic tools. The international bladder cancer nomogram provides a single 5-year recurrence estimate, but the models here offer 1-, 3-, and 5-year predictions for both recurrence and survival, enabling finer temporal risk stratification for individual patients.
Two key barriers must be addressed before clinical deployment. First, the models were trained at a single institution, and performance must be validated on multi-institutional datasets to ensure generalizability. Second, standardized collection of the 42-54 predictors used by the models must be ensured across clinical sites.
Adding genomic and molecular biomarker data to these clinicopathological predictors could further improve recurrence prediction accuracy, particularly for detecting metastatic disease that may be present but undetected at the time of surgery.
This work sets a benchmark for what is achievable using existing clinical data standards alone. In the absence of widely recognized biomarkers for bladder cancer, these models demonstrate the substantial predictive value locked within routine preoperative and operative data when analyzed with modern machine learning techniques.