3339 - Validation of Multi-Modal Artificial Intelligence Biomarker for Predicting Biochemical Recurrence in Patients with Prostate Cancer Undergoing Prostatectomy

A. Baydoun¹, Y. Sun², A. Y. Jia¹, N. G. Zaorsky¹, A. C. Callaway³, J. E. Shoag¹, R. A. Vince¹, L. Ponsky⁴, P. Mendiratta⁵, P. Barata⁵, J. Garcia¹, J. Brown¹, A. Berlin⁶, M. Ramotar⁶, A. Finelli⁷, C. J. D. Wallis⁸, T. Van der Kwast⁹, and D. E. Spratt¹⁰; ¹University Hospitals Seidman Cancer Center, Case Western Reserve University, Cleveland, OH, ²Case Western Reserve University School of Medicine, Cleveland, OH, ³University Hospitals Seidman Cancer Center, Cleveland, OH, ⁴University Hospitals Cleveland Medical Center, Case Western Reserve University School of Medicine, Cleveland, OH, ⁵UH Seidman Cancer Center, Case Western Reserve University, Cleveland, OH, ⁶Department of Radiation Oncology, Princess Margaret Cancer Centre, University of Toronto, Toronto, ON, Canada, ⁷Department of Surgical Oncology, Division of Urology, Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada, ⁸Mount Sinai Hospital, UHN, University of Toronto, Toronto, ON, Canada, ⁹University Health Network, Toronto, ON, Canada, ¹⁰Case Western, Cleveland, OH

Purpose/Objective(s): One potential advantage commonly cited for patients to undergo a radical prostatectomy (RP) is the ability to gain valuable prognostic information from the surgical pathology. However, this information is only available after a patient has underwent treatment. We previously developed the TRansfer learning-based multi-modal Artificial InteLligence model (TRAIL) for predicting 2 year-biochemical recurrence (BCR) following RP using artificial intelligence (AI)-derived features from digitized diagnostic biopsy slides. Herein, we aim to determine if we can achieve comparable prognostic performance of our pre-treatment TRAIL model to that of post-RP models that incorporate surgical histopathologic features. Materials/

Methods: Patients enrolled on a prospective registry were used for this study. Digital pathology slides from the diagnostic biopsy were scanned at 20x resolution. Cancerous regions within the slides were delineated using transfer learning on Inceptionv3 network that was trained on the PANDA dataset. Afterwards, direct supervision transfer learning was applied on AlexNet to extract 256 AI-derived features from cancerous pathology blocks. LASSO was used for AI features selection. TRAIL combines the clinical and selected AI features via a linear classification model. TRAIL was compared to two post-RP models; CAPRA-S, a purely clinicopathologic model trained to predict BCR post-RP, and a new model we created termed TRAIL-S which incorporates the post-RP data into the TRAIL model. The AI models were trained on 191 patients, then locked and evaluated on a separate testing set of 167 patients. TRAIL and TRAIL-S scores were then linearly transformed to the range of [0, 100]. Time dependent area under the curve (tdAUC) and C-index were used to evaluate model discrimination performance on the testing set.
Results: A total of 549 digital whole pathology slides from 358 patients who underwent RP were included. The median follow-up of the testing dataset is 49 months (95% CI: 45, 53). The pre-treatment TRAIL model had the highest tdAUC (0.83) & C-index (0.72) to discriminate 5-year BCR as compared to CAPRA (a pre-treatment non-AI model, tdAUC 0.72 & C-index 0.62), and both post-RP models (CAPRA-S tdAUC 0.82 & C-index 0.69, and TRAIL-S tdAUC 0.80 & C-index 0.62). When comparing patients below and above the median TRAIL score, patients with a low TRAIL score had an 80% relative reduction in risk of BCR compared to high TRAIL score patients (HR 0.20, 95% CI 0.08-0.51, p = < 0.001).
Conclusion: Our study demonstrates that the use of multi-modal AI from pre-treatment features alone to prognosticate BCR outcomes is equal or potentially superior than multivariable clinicopathologic models that contain post-RP prognostic information. This enables greater personalization pre-treatment to better understand each individual patients risk of recurrence prior to, rather than after, treatment.