Radical cystectomy is the standard treatment but outcomes vary widely. For patients with muscle-invasive bladder cancer (MIBC), radical cystectomy combined with neoadjuvant cisplatin-based chemotherapy in eligible patients is the standard of care. Yet despite apparently successful surgery, approximately half of all MIBC patients ultimately die from their disease due to occult micrometastases that were not detected preoperatively.
A fundamental problem is the discrepancy between clinical staging before surgery and final pathological staging of the removed specimen. Preoperative imaging and biopsy systematically underestimate tumor extent, and only the surgical pathology report provides definitive staging information. This means treatment decisions for neoadjuvant chemotherapy -- which must be given before surgery -- must be made with imperfect information.
Neoadjuvant chemotherapy is underutilized despite survival benefit. Despite evidence that neoadjuvant chemotherapy improves overall survival by approximately 5%, it is substantially underused globally. The modest absolute benefit combined with significant toxicity and the absence of reliable biomarkers to identify which patients will most benefit contributes to physician hesitancy. Better preoperative risk stratification tools could guide more targeted use of this treatment.
Several systemic inflammatory response (SIR) biomarkers had shown associations with adverse pathological features and cancer outcomes in prior single-biomarker studies, but none had been shown to meaningfully improve predictive models when combined with established clinicopathological variables. The hypothesis of this study was that a panel of SIR biomarkers evaluated together might achieve what individual biomarkers had not.
The inflammation-cancer connection. Systemic inflammatory response biomarkers reflect the balance between the immune system and tumor-promoting inflammation detectable through routine blood tests. They are derived from standard complete blood count and serum protein measurements, making them essentially free to obtain from blood samples already collected before surgery.
The five SIR biomarkers evaluated in this study were: the albumin-globulin ratio (AGR), which reflects protein balance and nutritional status; the neutrophil-lymphocyte ratio (NLR), measuring the balance between innate and adaptive immune activity; the monocyte-lymphocyte ratio (MLR); the De Ritis ratio (the ratio of liver enzymes AST/ALT used as a marker of cellular stress); and the modified Glasgow Prognostic Score (mGPS), which combines albumin and C-reactive protein levels into a three-point scale.
Why these markers might predict bladder cancer outcomes. High NLR and MLR indicate relative suppression of adaptive immunity with excess innate immune activity -- a state associated with tumor immune evasion. Low AGR reflects systemic protein depletion often seen in advanced cancer. The mGPS captures systemic inflammation through C-reactive protein combined with nutritional compromise through albumin. Together, these markers may capture different dimensions of the immune-tumor interaction that influence disease aggressiveness.
Prior individual studies of each biomarker in bladder cancer had shown associations with pathological outcomes and survival, but none had demonstrated that any single biomarker meaningfully improved prediction beyond what standard clinical variables already provide. Using all five markers together with machine learning to select the most informative combination represented a new approach to this question.
One of the largest bladder cancer surgical cohorts ever assembled. This retrospective multicenter study analyzed 4,199 patients who underwent radical cystectomy for clinically non-metastatic urothelial carcinoma from 12 international medical institutions between 1979 and 2012. All patients had preoperative serum measurements of all five SIR biomarkers taken within 4 weeks of surgery.
To simulate external validation, patients were randomly divided into a training cohort of 2,100 patients for model building and a testing cohort of 2,099 patients for independent performance evaluation. This split-sample approach tests whether patterns found in one half of the data replicate in a completely independent half, providing a more rigorous performance estimate than internal cross-validation alone.
Five clinical outcome endpoints. The study modeled three preoperative prediction targets: upstaging from clinically non-muscle-invasive to pathological MIBC at cystectomy (relevant for 1,527 patients staged cTis/cTa/cT1), lymph node involvement at cystectomy, and pT3/4 disease. Two postoperative survival endpoints were also modeled: cancer-specific survival (CSS) and recurrence-free survival (RFS). Patients who received neoadjuvant chemotherapy were excluded to isolate the biomarkers' independent predictive value.
Median follow-up for surviving patients was 41.9 months. Five-year RFS was 61.2%, five-year CSS was 67.3%, and five-year overall survival was 56.3%, reflecting the aggressive nature of muscle-invasive bladder cancer despite surgical treatment.
Addressing multicollinearity with regularization. Because SIR biomarkers share similar biological underpinnings and are expected to be correlated with each other, standard multivariable regression risks selecting redundant variables and producing unstable coefficient estimates. The study used LASSO (Least Absolute Shrinkage and Selection Operator) regression as a machine learning approach specifically suited to handle correlated predictors.
LASSO works by adding a penalty to the regression that forces less informative variables to have their coefficients shrunk to exactly zero, effectively excluding them from the model. This produces sparse, parsimonious models that contain only the most informative predictors. The strength of the penalty is controlled by a tuning parameter, with the optimal value determined by 10-fold cross-validation within the training cohort.
Building the most informative yet simplest model. Rather than selecting the LASSO penalty that minimized cross-validation error (which tends to include more variables), the study used the penalty that gave the simplest model within one standard deviation of the minimum error. This approach explicitly trades a small amount of predictive performance for substantially simpler models, which are more clinically practical and less prone to overfitting.
Calibration plots, goodness-of-fit tests, and 200-fold bootstrap correction of all performance metrics were applied to provide the most conservative and unbiased estimates of model performance. Decision-curve analysis was used to quantify the clinical net benefit of using the model versus treating all or no patients across a range of probability thresholds.
NLR was the most consistently selected biomarker. For all three preoperative prediction endpoints, the LASSO procedure selected at least one SIR biomarker as contributing unique predictive information. The neutrophil-lymphocyte ratio (NLR) was selected as the most effective predictor for both lymph node metastasis and upstaging to MIBC, with odds ratios of 2.26 and 1.97 respectively (both p less than 0.001), indicating that patients with higher NLR were approximately twice as likely to harbor these adverse features.
The bootstrap-corrected AUC for prediction of lymph node metastasis in the testing cohort was 67.3%. Prediction of pT3/4 disease reached AUC 73.0%, and prediction of upstaging to MIBC achieved AUC 65.8%. These figures represent the model's ability to rank patients by actual risk on a scale where 50% is chance and 100% is perfect discrimination.
Biomarkers added statistically significant but small gains. When SIR biomarkers were removed from the models and performance was re-evaluated with clinicopathological variables alone, the reduction in AUC was statistically significant for lymph node metastasis (5.8% decrease, p less than 0.001) and for upstaging prediction (4.3% decrease in testing cohort, p = 0.009). For pT3/4 disease, the decrease was below 3% and statistically significant only in the testing cohort.
Decision-curve analysis showed a meaningful clinical net benefit gain from adding SIR biomarkers for lymph node metastasis prediction at threshold probabilities between 30% and 50%, and for upstaging prediction at thresholds between 50% and 60%. These ranges contain the patients in whom the model's risk estimates are most likely to influence treatment decisions.
Different biomarkers selected for survival prediction. For postoperative cancer-specific survival, the LASSO process selected AGR, MLR, and De Ritis ratio as the most informative SIR biomarkers -- a different combination than for the preoperative staging models. For recurrence-free survival, AGR and MLR were selected. This divergence suggests that the biological mechanisms relevant to preoperative staging differ from those relevant to long-term survival after surgery.
The bootstrap-corrected C-index for cancer-specific survival in the testing cohort was 73.3%, and for recurrence-free survival was 72.2%. These C-index values reflect the models' ability to correctly rank pairs of patients by their actual survival outcomes -- a C-index of 50% would mean random ranking, and 100% would mean perfect discrimination.
Marginal improvement over clinicopathological variables alone. When SIR biomarkers were removed from the survival prediction models, the reduction in C-index was less than 3% for both CSS and RFS in both training and testing cohorts. Decision-curve analysis showed no relevant clinical net benefit gain from adding SIR biomarkers to reference models for either survival endpoint across any clinically meaningful probability threshold.
All models showed near-optimal calibration -- strong agreement between predicted probabilities and observed event rates across the full risk range. Time-dependent AUC analysis confirmed stable model performance over 5 years of follow-up, indicating that the models' discriminative ability did not degrade over time as patients' outcomes unfolded.
Statistical significance does not equal clinical utility. For a biomarker to be clinically useful, it must improve prediction by a meaningful margin -- generally considered an AUC or C-index increase of at least 5%, with the composite model exceeding 0.75. While several SIR biomarkers reached statistical significance in their incremental contribution, most improvements fell below 3%, which is below the threshold considered clinically meaningful for guiding individual patient decisions.
The authors propose that the underlying reason for this limitation is biological specificity: SIR biomarkers reflect general systemic inflammatory state but do not capture information that is uniquely relevant to bladder cancer biology. NLR, for example, is elevated in many inflammatory and oncological conditions and is not specific to the mechanisms driving bladder cancer invasion, lymph node spread, or metastasis.
The case for combination approaches. The study's negative findings point toward the need for biomarkers that capture more specific biological information. Genomic alterations detectable through next-generation sequencing hold greater biological specificity, but are hindered by cost, intratumoral heterogeneity, and lack of standardized external validation. The authors suggest that the highest potential lies in combining multiple data types: genetic, imaging, and blood-based biomarkers processed with artificial intelligence.
In the era of immunotherapy, SIR biomarkers may find a more targeted application. The inflammatory milieu captured by NLR, MLR, and mGPS is directly relevant to how tumors interact with the immune system, suggesting that these biomarkers might better predict responses to checkpoint inhibitors and other immunotherapies than they predict surgical outcomes.
Study limitations affect interpretation. The retrospective design spanning 1979 to 2012 introduces temporal heterogeneity -- surgical techniques, pathological assessment standards, and follow-up protocols all evolved significantly over this period. No patients received neoadjuvant chemotherapy, which reflects the underutilization of this treatment historically but limits applicability to current practice where more patients receive neoadjuvant therapy.
Several potential confounders were not controlled, including smoking history, prior intravesical treatments, and occupational exposures -- all of which influence bladder cancer biology and outcomes. While the multicenter design reduces single-institutional bias, it also introduces variation in lymph node dissection extent and urinary diversion choice that could affect outcomes.
A structured framework for future biomarker research. Despite negative results, the study establishes a rigorous methodological template for evaluating novel biomarkers in bladder cancer: LASSO-based variable selection to control for multicollinearity, split-sample validation to simulate external testing, 200-fold bootstrap correction for unbiased performance estimates, and decision-curve analysis to assess clinical net benefit beyond statistical significance alone.
The authors conclude that while SIR biomarkers remain potentially useful in the context of immunotherapy response prediction -- where the immune parameters they measure are more biologically relevant -- novel biomarkers with cancer-specific mechanistic links are still needed to meaningfully advance risk stratification and improve patient selection for perioperative systemic therapy in bladder cancer.