AUA 2024: Development and Validation of Generalizable Interpretable AI Biomarkers to Predict Clinical Outcomes in BCG-Treated Patients with Non-Muscle Invasive Bladder Cancer

(UroToday.com) The 2024 American Urological Association (AUA) annual meeting featured a session on artificial intelligence, and a presentation by Dr. Stephen Williams discussing development and validation of generalizable interpretable artificial intelligence biomarkers to predict clinical outcomes in BCG-treated patients with non-muscle invasive bladder cancer. Few markers exist that predict recurrence or progression in high-risk non-muscle invasive bladder cancer. Building clinically reliable artificial intelligence involves validating across major confounders, including (i) different cancer centers, (ii) heterogeneous patient risk factors and demographic subgroups, and (iii) imaging equipment variability. As such, Dr. Williams and colleagues built artificial intelligence to predict outcomes of high risk non-muscle invasive bladder cancer patients treated with BCG that generalized across:

  • 12 cancer centers in 4 continents
  • All clinically relevant subgroups within high risk non-muscle invasive bladder cancer
  • 8 different types of pathology scanners at 40x magnification

For this study, all patients underwent TURBT +/- re-TURBT and then intravesical BCG. Two digital H&E whole slide images from TURBT were used. Exclusion criteria included (i) inadequate pathologic tissue, (ii) induction/maintenance with non-BCG treatments post-TURBT, and (iii) muscle invasive, node positive, or metastatic disease. The development and validation of a computational artificial intelligence histology assay study, including 303 patients in the development cohort and 641 patients in the external validation cohort, is as follows:

With regards to bladder cancer stage, 34.1% were Ta, 54.8% T1, 11.1% CIS only, and 31.4% were any CIS. Adequate BCG (per the FDA definition) was given in 66% of patients, and 15.8% were BCG unresponsive:

image-1.jpg

Dr. Williams and colleagues took an interpretable approach to ensure that the artificial intelligence analyzes only clinically relevant feature in the pathology. The first step was to quantify the tumor microenvironment:

image-2.jpg

Board certified pathologists then created 300,000+ annotations of the tumor microenvironment to train the artificial intelligence algorithms, including:

  • Tumor features: nuclear atypia and hyperchromasia
  • Immune features: immune activation of lymphocytes and neutrophils
  • Stromal features: vasculature and fibroblasts

Of note, 62 million micron2 of tissue was used to train the algorithms. Artificial intelligence quantification of tumor cells, stromal cells, and immune cells was consistent across many cell types in the development cohort with an AUC of > 0.94:

image-3.jpg

The computational artificial intelligence histology assay maintained strong concordance on predictions across all major digital pathology scanner models. This is important given that variations in scanner type can render artificial intelligence models unusable in the real world. This artificial intelligence model was validated across all major scanner brands (Phillips, Leica, Hamamatsu, and 3D Histech) to demonstrate consistent performance, including a subset of slides that were scanned multiple times across each major scanner brand (correlation between scanner types R2 > 0.95). 

image-4.jpg

The second step was using the tumor microenvironment features to identify a signature associated with outcomes in the development cohort:

image-5.jpg

Digital features associated with high-grade recurrence free survival and progression free survival were identified in the development set and then feature-locked biomarker assays were tested on an independent validation set:

image-6.jpg

For recurrence, the artificial intelligence model independently prognosticates high grade recurrence (HR 2.0, p < 0.001):

image-7.jpg

For progression, the artificial intelligence model independently prognosticates progression (HR 3.2, p < 0.001):

image-8.jpg

When comparing the computational artificial intelligence histology assay versus EORTC 2016 for stratifying recurrence free survival in the validation cohort, at 2 years patients with an AI-high risk of recurrence had a 2x higher risk of high grade recurrence free survival post-BCG (HR 2.1, 95% CI 1.8-2.4, p < 0.001). EORTC shows unstable prognostication with crossover of group survival plots: 

image-9.jpg

When comparing the computational artificial intelligence histology assay versus EAU NMIBC 2021 progression risk scoring for stratifying muscle invasive progression in the validation cohort, at 3 years patients with an AI-high risk of progression had a 3x higher risk of progression post-BCG (HR 3.9, 95% CI 2.8-5.4, p < 0.001). This is in comparison to the EAU “very high risk” vs “high risk” having a hazard ratio of only 1.51 (p < 0.03):

image-10.jpg

Finally, the computational artificial intelligence histology assay performs across the heterogeneity of high risk NMIBC by maintaining significance across all clinical subgroups:

image-11.jpg

Dr. Williams concluded his presentation by discussing the development and validation of generalizable interpretable artificial intelligence biomarkers to predict clinical outcomes in BCG-treated patients with non-muscle invasive bladder cancer with the following take-home messages:

  • This computational histology artificial intelligence assay is undergoing an Early Access Program as a CLIA-approved lab developed test with leading centers
  • The assay maintains generalizability across different digital slide scanners and across patient subgroups
  • By understanding an individual’s risk of recurrence, progression, cystectomy, and developing BCG unresponsive disease, the artificial-based multimodality assay:
    • Arms clinicians with nuanced data to help guide treatment selection and patient counseling regarding alternative treatment approaches
    • Better places clinicians for clinical decision-making during an era of BCG shortage and when novel therapies or therapeutic combinations become available since this assay could help identify patients less likely to benefit from BCG
    • This artificial intelligence model may have implications for tumor surveillance regimens 

Presented by Stephen B. Williams, MD, MS, FACS, University of Texas Medical Branch, Galveston, TX

Written by: Zachary Klaassen, MD, MSc – Urologic Oncologist, Associate Professor of Urology, Georgia Cancer Center, Wellstar MCG Health, @zklaassen_md on Twitter during the 2024 American Urological Association (AUA) Annual Meeting, San Antonio, TX, Fri, May 3 – Mon, May 6, 2024.