Development and validation of a computational histology artificial intelligence (CHAI) prognostic biomarker in renal cell carcinoma.

Abstract

546 Background: Artificial intelligence (AI)–based digital pathology has shown great promise in prostate and bladder cancer, but its application in renal cell carcinoma (RCC) remains limited. We applied the Computational Histology Artificial Intelligence (CHAI) platform, a deep learning-based pathology system, to develop and validate a novel prognostic biomarker for localized RCC. Methods: Hematoxylin and eosin stained whole-slide images from stage I-III RCC cases in The Cancer Genome Atlas (TCGA) were analyzed with CHAI to quantitatively extract histologic features from the tumor and microenvironment. All RCC subtypes within TCGA were included (clear cell, papillary, chromophobe). Cases were divided into development (30%) and validation (70%) cohorts using a stratified random split. In the development cohort, a continuous histologic risk signature for overall survival (OS) was derived and dichotomized into low- and high- risk groups with a 70-30 cutoff. We then tested the locked biomarker and evaluated its performance in the independent validation cohort. Multivariate Cox proportional hazards (CPH) models assessed associations with progression-free survival (PFS), disease-specific survival (DSS), and OS. Results: 801 RCC patients were available (236 in development, 565 in validation). In the validation cohort, the median age was 60 years (IQR 51-70), and 370 (65%) were male. Disease stage was I in 340 (60%), II in 86 (15%), and III in 139 (25%). Histologic subtypes in the validation cohort included clear cell in 307 (54%), papillary in 184 (33%), and chromophobe in 74 (13%). On validation, the biomarker stratified patients into 167 (29%) high-risk and 398 (70%) low-risk. High-risk status was significantly associated with inferior PFS (HR: 2.74 [95%CI: 1.81, 4.15]), DSS (HR: 3.37 [1.89, 6.03]) and OS (HR: 1.92 [1.25, 2.96]), all p<0.05. Associations remained significant after controlling for age, sex, stage, and subtype (p<0.05). At 1 year, high-risk patients demonstrated higher rates of progression (15% vs 4.6%), disease-specific mortality (4.8% vs 1.1%), and overall mortality (7.8% vs 2.9%). Conclusions: We developed and validated an AI-derived digital pathologic biomarker that prognosticates localized RCC by risk of progression, disease-specific death, and overall mortality--independent of histologic subtype. These findings support further investigation of CHAI-based digital biomarkers as tools to refine RCC risk assessment and guide personalized management. Multivariate CPH model of OS. Variable Factor HR (95% CI) p-value Biomarker (ref: Low-Risk) High-Risk 1.90 (1.22, 2.97) <0.01 Age 1.03 (1.01, 1.04) <0.01 Sex (ref: Female) Male 0.71 (0.45, 1.11) 0.12 Stage (ref: Stage I) II 1.61 (0.79, 3.29) 0.19 III 3.93 (2.43, 6.35) <0.01 Subtype (ref: Clear Cell) Papillary 0.74 (0.43, 1.25) 0.26 Chromophobe 0.43 (0.18, 1.02) 0.06