Predicting Metastatic Progression of High Risk Localized Prostate Cancer

点击翻译按钮获取中文摘要

ABSTRACT. Prostate cancer (CaP) is the most commonly diagnosed malignancy other than non-melanoma skin cancer amongst Veterans. Approximately 7% of US CaP cases are diagnosed and treated in the Veteran population. High risk (HR), localized CaP represents 20-25% of the approximately 250,000 new cases of CaP expected in the US in 2022. The outcomes of HR CaP are variable, with some patients remaining in remission and others suffering from metastatic progression and death. Our ability to discriminate between patients who will fare well following curative-intent treatment versus those destined for lethal metastatic progression remains poor. Our overall objective is to apply artificial intelligence (AI) algorithms to generate novel predictors of metastasis-free survival (MFS), the only validated surrogate for overall survival in localized CaP, from a large repository of digital pathology and radiographic images. We will then combine these AI-derived biomarkers with clinical-pathologic and social determinants of health (SDoH) variables collected from Veterans with HR CaP to develop and test multivariable prognostic models that improve our ability to predict MFS. AI, including computer vision and machine learning approaches, allows extraction of image patterns for sub- visual based characterization of CaP. Routine diagnostic prostate needle biopsy pathology slides that have been digitized as well as digital radiographic images (e.g. MRI) can be leveraged for machine learning derived from either (1) hand-crafted features (guided by existing domain knowledge) which are then used as the inputs to develop the machine-learning model based on the selected features, or (2) the raw data itself, which are used as inputs to develop the model through convolutional neural networks or other methods in an unsupervised manner. The former leverages existing domain knowledge and may require less input data, whereas the latter is not limited by prior knowledge, but requires more training data. We hypothesize that machine learning models based on multimodal data derived from MRI and digital pathology can be combined with clinic-pathologic and SDoH data to generate “super classifiers” that more accurately predict outcome without the need for costly tissue destructive methods. We propose to establish a collection of digital pathology and prostate MRI images along with clinic-pathologic and SDoH data from >5,000 Veterans with HR CaP who have been treated with curative intent and a minimum of 5 years of follow-up using our existing approved biorepository protocol. Subsequently, we will determine the most robust AI algorithm for each data source, and then test combinations of algorithms to generate a “superclassifier” that integrates AI-derived predictive models with standard clinico-pathologic and SDoH variables to predict MFS. Improved prognostication could illuminate strategies for treatment intensification or de- intensification that can be formally tested in future clinical trials. The substantial infrastructure and databases generated by this proposal as part of our repository will be accessible by intramural VA and extramural investigators for future approved studies.