Jun Deng

Yale University

Professor

PLCO (Learn more about this study)

PLCO-365

May 16, 2018

Stratifying Prostate Cancer Risk based on Deep Learning of PLCO Data

Prostate cancer is the most common cancer among men except for skin cancer, and is the third leading cause of cancer death in men in the United States. It is estimated that there are about 161,360 new cases in America each year. Although the 5-year survival rate for the localized and regional prostate cancer (often low-risk) is about 100%, it is only 29% for the metastatic and advanced cases (high-risk prostate cancer). Despite a tremendous amount of money and resources have been spent on prostate cancer screening and treatment over the years, more than 26,730 men die from prostate cancer each year in the United States, mostly from high-risk prostate cancer. It is hence highly significant and desirable to have a screening method that can distinguish high-risk prostate cancer from low-risk one for more precise intervention and prevention. Over the past few years, artificial intelligence has shown great promise in leveraging big health data to diagnose and stage diseases, reduce cost, improve health care and patient outcome. While access to big data stored in silo-like electronic medical record systems has been very restrictive, the large amount of health data collected in some nation-wide multi-centered studies, such as the National Health Interview Survey (NHIS) and the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO), is readily available. Recently, we have demonstrated that a deep neural network (DNN), trained and validated with the NHIS datasets of 1997 to 2016, can be used for prostate cancer prediction based only on personal health data, without family history, cancer staging or follow-up information. Our DNN has achieved a sensitivity of 21.5%, specificity of 91%, positive predictive value (PPV) of 28.5%, and the area under of curve (AUC) of 0.73. Built upon our preliminary study, we hypothesize that the large amount of data collected in the PLCO trial, including not only personal health data but also the staging of prostate cancer, long-term follow-up, family history, socio-behavioral and lifestyle data, dietary and PSA data, can be used to train and validate a deep learning algorithm to stratify prostate cancer risk for individuals. Hence, the goal of this project is to develop a risk predictor and classifier based on deep learning of PLCO data to distinguish high-risk prostate cancer from low-risk prostate cancer for better risk stratification and more precise intervention, hence improving outcomes for high-risk patients while minimizing overtreatment of low-risk disease. With the daunting cost of healthcare and nontrivial rate of cancer mortality, it is critical to stratify an individuals prostate cancer risk prior to its onset to maximize outcomes for patients with aggressive disease and minimize unnecessary treatment of indolent disease. We believe this project is a step toward achieving this goal by focusing on prostate cancer risk prediction and classification at individual level. If implemented successfully, we envision a risk predictor and classifier tool deployed in the clinic for effective prostate cancer risk stratification and prevention for millions of people worldwide, hence reducing prostate cancer mortality in the long run.