Validation of a computational histology artificial intelligence (CHAI) prognostic biomarker in muscle invasive bladder cancer.

Abstract

791 Background: Standard management for MIBC is radical cystectomy (RC). This traditionally includes neoadjuvant chemotherapy (NAC) which confers additional survival benefit. However, the absolute survival benefit of NAC is 5-10%, and only 30-40% of US MIBC cases receive NAC. These limitations underscore the need for a more personalized approach to identify patients most likely to benefit from treatment escalation or avoid overtreatment in those likely cured with RC alone. The CHAI biomarker (BM) platform has previously been used to develop a now commercially-available prognostic tool in non-MIBC. We aimed to validate a novel prognostic BM, previously developed specifically for MIBC using the CHAI platform, in a multi-institutional real-world (RWD) MIBC cohort. Methods: The CHAI platform applies deep-learning to extract quantitative histologic features from pre-treatment transurethral resection bladder tumor (TURBT) specimen H&E-stained whole-slide images. The previously-developed MIBC BM was locked; the model outputs a continuous histologic risk signature and is dichotomized into unfavorable BM positive (BM+) and favorable BM negative (BM-) groups. Validation was done on an independent, held-out, retrospective, pooled, RWD of clinical T2N0M0 MIBC patients who underwent RC at multiple NCI-Designated Centers. Multivariate (MVA) Cox proportional hazards (CPH) models assessed associations with recurrence-free survival (RFS), cause-specific survival (CSS), and OS. Kaplan-Meier methods and log-rank tests were used for survival analysis. Results: Among 134 patients with MIBC, 67 (50%) patients received cisplatin-based NAC. The BM stratified patients into 67 (50%) BM+ and 67 (50%) BM-. Among patients who received NAC, pathologic complete response (pCR) was seen at RC among 6/29 (20%) BM+ vs 13/38 (34%) BM-. On MVA, controlling for age, sex, presence of carcinoma in situ (CIS), variant histology, and NAC, BM+ was significantly associated with inferior RFS (Table; HR: 2.70 [1.46, 4.97]), CSS (HR: 2.60 [1.29, 5.26]) and OS (2.29 [1.29, 4.04]), all p<0.01. With 36-mo median follow up, at 2yr, BM+ had worse outcomes: RFS 50% vs 18%; CSS 25% vs 8.3%; OS 36% vs 15%. Conclusions: A histologic prognostic BM derived from pre-treatment H&E TURBT slides was validated, stratifying clinical T2 MIBC patients by risk of RFS, CSS, and OS, even when controlling for NAC. Future efforts will evaluate the ability to predict which patients benefit the most from additional treatment. Such a tool could be used to optimize patient selection by identifying those most likely to benefit from perioperative therapies, and sparing those who may be cured with surgery alone. MVA for RFS. HR (95% CI) P value Biomarker 2.70 [1.46, 4.97] p=0.01* Age 1.03 [1.00, 1.07] p=0.06 Sex 1.24 [0.65, 2.35] p=0.52 CIS 0.69 [0.32, 1.50] p=0.35 Variant histology 1.15 [0.56, 2.35] p=0.71 NAC 1.54 [0.80, 2.95] p=0.20