Prophet: Deep Learning to Predict the Early Onset of Cancers
Principal Investigator
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCOI-505
Initial CDAS Request Approval
Aug 8, 2019
Title
Prophet: Deep Learning to Predict the Early Onset of Cancers
Summary
1 in 70 women will get ovarian cancer over her lifetime, more than half will not survive longer than 5 years. This is mostly due to the fact that 70% of these cases are diagnosed at a stage later than stage 2. Currently, there is no highly reliable early detection method for screening for ovarian cancer.
By using data from the PLCO and Deep Learning, we estimate that we can improve the accuracy of ovarian cancer screenings. We aim to use a combination of imaging data, tabular data, and patient demographics to predict a patient's risk rating, guiding them through to a secondary screening/biopsy.
We will use similar methods that we used for researching signs of the early prevalence of cervical cancer. We will first gather and aggregate data by patient identifiers, creating "pseudo-records" using data from the NIH. Then, we will process each "p-record" through our Adaptive Recurrent Convolutional Neural Networks (ARCNNs), a novel approach for dealing with diverse datasets.
We will then create risk-based summary statistics like match ratings, class distributions, risk ratings, confidence, etc. We aren't testing a particular hypothesis, we are testing multiple hypotheses against one another, drawing correlations, and optimizing for those that hold true through multiple generations.
By using data from the PLCO and Deep Learning, we estimate that we can improve the accuracy of ovarian cancer screenings. We aim to use a combination of imaging data, tabular data, and patient demographics to predict a patient's risk rating, guiding them through to a secondary screening/biopsy.
We will use similar methods that we used for researching signs of the early prevalence of cervical cancer. We will first gather and aggregate data by patient identifiers, creating "pseudo-records" using data from the NIH. Then, we will process each "p-record" through our Adaptive Recurrent Convolutional Neural Networks (ARCNNs), a novel approach for dealing with diverse datasets.
We will then create risk-based summary statistics like match ratings, class distributions, risk ratings, confidence, etc. We aren't testing a particular hypothesis, we are testing multiple hypotheses against one another, drawing correlations, and optimizing for those that hold true through multiple generations.
Aims
Explore as much data related to ovarian cancer as possible.
Develop correlations related to screening methods, external factors, and cyst factors to ovarian cancer.
Build a model to reliably predict the risk of a patient developing ovarian cancer.
Achieve a target, individual, PPV of over 10%. We estimate that a combined (CA-125, ultrasonography, and our model) PPV could exceed 40% (compared to the gold standard of 26.8%).
Collaborators
Dr. Shahnoz Rustamova - Central Park Medical Practice
Dr. Nimesh Nagarsheth - Mt. Sinai