AN ARTIFICIAL INTELLIGENCE-POWERED MODEL TO PREDICT RESPONSE TO INTRAVESICAL BCG VERSUS GEMCITABINE-DOCETAXEL FOR HIGH-RISK NON-MUSCLE INVASIVE BLADDER CANCER

Vignesh T. Packiam; Ian McElree; Jordan Richards; Anand Rajan; Stephanie Chen; Viswesh Krishna; Snehal Sonawane; Vivek Nimgaonkar; Damir Vrabac; Waleed Abuzeid; Anirudh Joshi; Ashish Kamat; Drew Watson; Michael O'Donnell

doi:10.1016/j.urolonc.2024.12.209

Intravesical BCG is first-line adjuvant therapy for high-risk non-muscle invasive bladder cancer (HR-NMIBC). Sequential intravesical gemcitabine and docetaxel (Gem/Doce) has increasingly shown promising results as therapy in the BCG-naive setting. Recently, we demonstrated a computational histology artificial intelligence (CHAI) assay that predicts response to BCG using digital images of H&E pathology slides from pre-treatment transurethral resection of bladder tumor (TURBT) specimens. Presence of the biomarker conferred an increased risk of recurrence, progression, and BCG unresponsive disease (BUD). We aimed to assess if this biomarker was predictive of response to BCG versus Gem/Doce. Pre-treatment TURBT-derived digital whole slide images (WSI) and clinical data were obtained for patients with BCG-naïve AUA HR-NMIBC diagnosed between 2011-2021 and treated with BCG or Gem/Doce at the University of Iowa. WSIs were analyzed using the CHAI assay incorporating histologic features previously shown to be predictive of development of BUD. Those with “biomarker present” were previously shown to have a shorter time to BUD than those with “biomarker absent”. High-grade recurrence-free survival (HG-RFS) between BCG and Gem/Doce were compared in the pre-defined biomarker present and absent groups. A test for statistical interaction between the biomarker and treatment type was assessed via Cox proportional hazards regression and likelihood ratio tests. There were 184 patients with HR-NMIBC, of whom 105 received BCG and 79 received Gem/Doce (31% HGTa, 59% HGT1, 35% with any CIS). Median follow-up was 36 months. Among patients who received BCG, “biomarker present” cases had inferior HG-RFS compared to “biomarker absent” cases (p=0.027). Biomarker presence was detected in 60 patients (33%) of whom 33 received BCG and 27 received Gem/Doce. “Biomarker present” cases treated with Gem/Doce had significantly higher HG-RFS compared to those treated with BCG (HR 0.13, 95% CI [0.06, 0.28], p<0.001) (Figure 1A). In the 124 “biomarker absent” cases (67%), there was no significant difference in HG-RFS between Gem/Doce and BCG (HR 0.64, 95% CI [0.42, 0.99], p=0.3) (Figure 1B). In the “biomarker present” cases, the HG-RFS when treated with Gem/Doce was 93.6% compared to only 54.6% when treated with BCG (logrank p=0.0019). The interaction test for biomarker and treatment type was significant (p<0.05), suggesting that the biomarker is predictive of response to intravesical BCG and GD. This AI biomarker is one of the first predictive tests with the potential to guide adjuvant intravesical therapy for HR-NMIBC. The CHAI biomarker assay not only identifies patients with inferior response to BCG but also facilitates selection of patients more likely to respond to Gem/Doce. Larger validation studies are ongoing.

AN ARTIFICIAL INTELLIGENCE-POWERED MODEL TO PREDICT RESPONSE TO INTRAVESICAL BCG VERSUS GEMCITABINE-DOCETAXEL FOR HIGH-RISK NON-MUSCLE INVASIVE BLADDER CANCER

View Online

Abstract

Details

Metrics