Pd27-12 Development and Validation of Generalizable Interpretable AI Biomarkers to Predict Clinical Outcomes in BCG-Treated Patients With Non-Muscle Invasive Bladder Cancer

You have accessJournal of UrologySurgical Technology & Simulation: Artificial Intelligence II (PD27)1 May 2024PD27-12 DEVELOPMENT AND VALIDATION OF GENERALIZABLE INTERPRETABLE AI BIOMARKERS TO PREDICT CLINICAL OUTCOMES IN BCG-TREATED PATIENTS WITH NON-MUSCLE INVASIVE BLADDER CANCER Yair Lotan, Jay B. Shah, Viswesh Krishna, Bryn Launer, Vrishab Krishna, Siddhant Shingi, Jennifer Gordetsky, Thomas Gerald, Eugene Shkolyar, Dickon Hayne, Andrew Redfern, Lisa Spalding, Courtney Stewart, Vikram Narayanan, Dattatreya Patil, Vignesh T. Packiam, Anand Rajan, Michael A. O'Donnell, Loic Baekelandt, Mario I. Fernandez, Marcela Schultz, Patrick J. Hensley, Derek B. Allison, John A. Taylor, Ameer Hamza, Vivek Nimgaonkar, Ekin Tiu, Louis J. Vaickus, Snehal Sonawane, Daniel L. Miller, Damir Vrabac, Waleed M. Abuzeid, Anirudh Joshi, Sam S. Chang, and Stephen B. Williams Yair LotanYair Lotan , Jay B. ShahJay B. Shah , Viswesh KrishnaViswesh Krishna , Bryn LaunerBryn Launer , Vrishab KrishnaVrishab Krishna , Siddhant ShingiSiddhant Shingi , Jennifer GordetskyJennifer Gordetsky , Thomas GeraldThomas Gerald , Eugene ShkolyarEugene Shkolyar , Dickon HayneDickon Hayne , Andrew RedfernAndrew Redfern , Lisa SpaldingLisa Spalding , Courtney StewartCourtney Stewart , Vikram NarayananVikram Narayanan , Dattatreya PatilDattatreya Patil , Vignesh T. PackiamVignesh T. Packiam , Anand RajanAnand Rajan , Michael A. O'DonnellMichael A. O'Donnell , Loic BaekelandtLoic Baekelandt , Mario I. FernandezMario I. Fernandez , Marcela SchultzMarcela Schultz , Patrick J. HensleyPatrick J. Hensley , Derek B. AllisonDerek B. Allison , John A. TaylorJohn A. Taylor , Ameer HamzaAmeer Hamza , Vivek NimgaonkarVivek Nimgaonkar , Ekin TiuEkin Tiu , Louis J. VaickusLouis J. Vaickus , Snehal SonawaneSnehal Sonawane , Daniel L. MillerDaniel L. Miller , Damir VrabacDamir Vrabac , Waleed M. AbuzeidWaleed M. Abuzeid , Anirudh JoshiAnirudh Joshi , Sam S. ChangSam S. Chang , and Stephen B. WilliamsStephen B. Williams View All Author Informationhttps://doi.org/10.1097/01.JU.0001008580.58088.27.12AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Few markers exist that predict recurrence or progression in high-risk non-muscle invasive bladder cancer (HR NMIBC). We developed a deep learning pipeline that extracts generalizable interpretable features from digitized pathology slides. We validated histologic assays that predict recurrence, progression and BCG response in a multicenter cohort. METHODS: Digital H&E pathology slides and longitudinal clinical data were collected. Existing pretreatment diagnostic slides were converted to whole slide images and were not recut or re-stained to optimize model generalizability. Cell and tissue segmentation and classification models were developed from >1.5 million data points annotated by pathologists. Digital features associated with high-grade recurrence free survival (RFS), progression free survival (PFS) and BCG response were identified in the development (dev) set and then feature-locked biomarker assays were tested on an independent validation (val) set. Patients were classified into high (AI-H) and low (AI-L) risk of recurrence and progression and for presence/absence of a BCG Unresponsive Biomarker (BUB). RESULTS: Cell and tissue models had 0.99 AUC on the test set. AI models generalized across centers & slide preparation (Figure 1) and were robust to scanner model & magnification (r=0.99). 1071 HR NMIBC patients (dev: 311, val: 760) from 12 centers (med followup: 34 mo) were included. Features related to tumor aggressiveness, nuclear atypia, and immune response were associated with clinical outcomes. On regression analysis, AI groups predicted outcomes independent of age, sex, smoking status, grade, stage, presence of CIS and multifocality. In the val set, AI-H cases had significantly inferior RFS and PFS vs AI-L cases (HR 2.32, 3.08, p<0.001) with higher recurrence risk (1.9x & 1.7x at 1 & 2 yrs) and progression risk (2.9x & 2.5x at 1 & 5 yrs). BCG unresponsive status was 1.8x higher in BUB+ vs BUB- cases. CONCLUSIONS: We developed and validated AI assays that are robust to real-world variations across centers, preparation and digital scanning methods. These assays can be used to identify HR NMIBC cases with significantly higher risk of recurrence, progression, and BCG unresponsive status who may benefit from alternative therapies. Download PPT Source of Funding: None © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e555 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Yair Lotan More articles by this author Jay B. Shah More articles by this author Viswesh Krishna More articles by this author Bryn Launer More articles by this author Vrishab Krishna More articles by this author Siddhant Shingi More articles by this author Jennifer Gordetsky More articles by this author Thomas Gerald More articles by this author Eugene Shkolyar More articles by this author Dickon Hayne More articles by this author Andrew Redfern More articles by this author Lisa Spalding More articles by this author Courtney Stewart More articles by this author Vikram Narayanan More articles by this author Dattatreya Patil More articles by this author Vignesh T. Packiam More articles by this author Anand Rajan More articles by this author Michael A. O'Donnell More articles by this author Loic Baekelandt More articles by this author Mario I. Fernandez More articles by this author Marcela Schultz More articles by this author Patrick J. Hensley More articles by this author Derek B. Allison More articles by this author John A. Taylor More articles by this author Ameer Hamza More articles by this author Vivek Nimgaonkar More articles by this author Ekin Tiu More articles by this author Louis J. Vaickus More articles by this author Snehal Sonawane More articles by this author Daniel L. Miller More articles by this author Damir Vrabac More articles by this author Waleed M. Abuzeid More articles by this author Anirudh Joshi More articles by this author Sam S. Chang More articles by this author Stephen B. Williams More articles by this author Expand All Advertisement PDF downloadLoading ...