Background
Accurate prognosis is a cornerstone of high-quality cancer care. It helps clinicians tailor treatments to patient’s needs — maximizing effectiveness while minimizing unnecessary harm — and gives patients realistic expectations about their disease trajectory. In bladder cancer, this need is particularly urgent. Recurrence rates are among the highest of all cancers, making precise prognostic assessment critical.
For decades, the UICC staging system has served as a global standard in cancer classification — enabling consistent patient stratification and guiding treatment decisions across healthcare systems Yet in clinical practice, patients within the same UICC stage can have vastly different disease trajectories. Some progress rapidly, while others remain stable for years. This variability highlights a key limitation of anatomy-based staging: it does not sufficiently capture the biological behavior of the tumor. To enhance prognostic precision and better tailor treatment decisions, we need additional tools that better account for the biological heterogeneity of each patient's tumor.
One promising approach lies in the tumor microenvironment (TME), which plays a crucial role in cancer progression. However, analyzing the TMEat scale remains challenging due to reliance on specialized staining techniques, time-intensive manual assessment, and lack of standardized computational tools suitable for routine clinical use.
Our study presented at this year's ASCO meeting addressed this challenge by leveraging our Atlas H&E-TME profiling tool. Using standard H&E-stained slides, we explored whether combining TME features with conventional UICC staging could improve prognostic stratification in bladder cancer.
Methods
In our study, we analyzed a bicentric cohort of over 700 patients with resected bladder cancer. All patients had undergone radical cystectomy and were followed for a median of 28 months to track clinical outcomes. The cohort spanned the full range of disease stages, from UICC stage I to IV.
TME feature analysis was conducted using our Atlas H&E-TME tool, which performs comprehensive analysis of the TME at single-cell resolution in H&E images. Atlas H&E TME performs slide QC, tissue segmentation, cell detection, and cell classification to generate 5000+ readouts and polygon overlays for each whole slide image.
Of the readouts, 26 spatially resolved cell densities were selected and then integrated with clinicopathological variables – such as the UICC stage - to determine if details about the TME could improve prognostic stratification.
Results
To determine the impact of TME features on prognostic stratification, we built and tested Cox proportional hazards regression models to analyze the influence of input variables on survival. In the baseline model, we used the UICC stage alone to predict outcomes — reflecting current clinical practice. In the extended model, we combined UICC staging with selected TME features to determine whether the integration of TME data could improve predictive performance.
From the combined UICC + TME model, we calculated patient risk scores, which give a predicted mortality risk for each patient. We then grouped the patients into four equi-distant risk groups based on these scores and evaluated how well this new stratification performed compared to the original UICC staging groups using three analytical approaches.
How accurate are the model's predictions?
To validate that our Cox regression models reliably rank patients by risk, we calculated the C-index (concordance index), where 0.5 = random performance and 1.0 = perfect prediction. This calculation was done multiple times over 25 combinations of data splits and initializations to ensure statistical significance. We found that:
- UICC-only model: C-index of 0.611 ± 0.035
- UICC + TME model: C-index of 0.627 ± 0.038 (p < 0.001)
The higher C-index and significant p-value show that adding TME features creates a model that is more accurate at separating high-risk from low-risk patients.
How large is the difference between risk groups?
To quantify how mortality risk differs between patient groups, we calculated hazard ratios, which compare the risk of an event at any given time between groups. This analysis compared binary groups split by median risk: predicted high-risk versus predicted low-risk. Results showed:
- UICC-only model: Hazard ratio of 1.749 (95% CI: 1.451 - 2.109), p-value = 4.539e-09 - meaning high-risk patients had about a 75% higher chance of dying at any given moment compared to low-risk patients
- UICC + TME model: Hazard ratio of 1.971 (95% CI: 1.635 - 2.378), p-value = 1.243e-12 - meaning high-risk patients had about a 97% higher chance of dying at any given moment compared to low-risk patients
The much stronger hazard ratio and highly significant p-value shows that adding TME features creates better differentiation between high and low-risk groups.
How well does the model stratify risk groups?
To determine if combining TME features with UICC data improves patient stratification, we calculated and modeled Kaplan-Meier plots, which visualize cumulative survival probability over time. In the example figure below, we specifically looked at patients in UICC stage III as it had the least clarity when compared to the UICC + TME model.
While the UICC model classified all patients as Stage III, our TME-enhanced model successfully stratified this population into distinct high-risk and low-risk groups. High-risk patients (yellow curve) demonstrated consistently lower survival probabilities with steeper decline over time compared to low-risk patients (blue curve). This demonstrates how patients within the same UICC stage can have vastly different outcomes. Our TME-based approach revealed these hidden differences, offering a more refined way to stratify risk within stage III and beyond.
Conclusion
Our TME-enhanced model outperformed traditional UICC staging, achieving higher predictive accuracy and clearer separation of prognostic risk groups. This demonstrates that integrating TME data from routine H&E slides with UICC staging improves risk stratification, helping pinpoint high-risk bladder cancer patients more precisely than anatomical staging alone. By using only standard H&E staining available in all clinical settings, this approach of extracting TME data is more accessible than methods requiring specialized staining techniques, while still providing enhanced prognostic accuracy.
In the future, we will validate our findings across larger, multicenter cohorts, investigate spatially-resolved cellular neighborhoods and assess whether TME features also hold predictive value for treatment response. By expanding both the scale and scope of our analysis, we aim to demonstrate that this approach can systematically improve patient stratification and enhance clinical decision-making in bladder cancer care.