Study
PLCO
(Learn more about this study)
Project ID
PLCO-786
Initial CDAS Request Approval
May 11, 2021
Title
AI system to infer a patient’s prostate cancer stage from clinical information trained on synthetic electronic health records
Summary
Clinical staging of prostate cancer is the first important step in determining the treatment course for a patient with a confirmed prostate cancer diagnosis. Staging work-up findings such as PSA, tumor grade, and genetic subtype could objectively determine the staging and lead to an optimal treatment recommendation according to for example NCCN guidelines. However, in practice, clinical tumor staging determination is a subjective; different multidisciplinary team clinicians may assign different clinical stages to the same patient. Worse, pertinent findings from the patient history and physical examination and radiographic interpretations are usually recorded in unrestricted clinical text that makes retrieving these pieces of information difficult and error prone. It is highly desirable to have a system that could reduce the burden of data extraction and aggregation and could infer based on given rules an accurate and consistent clinical staging for a given patient. To this end, we propose a novel artificial intelligence based system that could accurately and automatically extract pertinent information from the medical record, and through a trained network infer the clinical stage. We specifically address the limited training data access due to privacy restrictions, by creating synthetic patients charts reflecting expected range of clinical parameters used for the staging process.
Aims
SA1: Identify variability in diagnostic testing and results for patients suspicious of having prostate cancer. This include variabilities in sequence of diagnostics tests, and specific set of results for each test. This step could be based on combination of real-data, or cohort level statistics presented in various publications.
SA2: Build a synthetic patient chart generator taking into account possible variabilities learnt during SA1. Furthermore, the synthetic patients charts should includes both coding and description variations for detailed of the test results with possible inclusion of inadvertent errors and missing information. The synthetic patient chart must also include the correct clinical staging information assigned at the end of the staging work-up.
SA3: Build an AI system to infer clinical staging information based a set of diagnostic tests and corresponding results from a patient chart. The system is trained based on a combination of synthetic and real data with known clinical stage information
Collaborators
