Deep Learning for Pulmonary Embolism Imaging Decision Support: A Multi-institutional Collaboration

项目总结

Project Summary Diagnostic imaging costs $100 billion annually. These healthcare costs are expected to increase in the coming decade as the national population ages and the pool of insured patients increases. The size and growth of these costs concern policy makers, payers, and society alike. The use of advanced imaging for PE has increased 27 fold in recent years, and this sharp escalation has the potential to expose patients to unnecessary procedures, tests, and risks due to incidental findings. Although radiologists do not order most radiology exams, these physicians are the target of criticism about the rising costs and possible overuse of radiology services. The healthcare industry has called upon radiologists to manage the potential overuse of advanced imaging and to take the lead on investigating best practices for the optimal use of advanced imaging. The ideal sources of information for imaging utilization guidelines are randomized, controlled imaging clinical trials. However, these trials are cost and time intensive, exceedingly difficult to conduct, and typically use narrow patient-inclusion criteria, making it challenging to generalize the results to broader clinical situations. Alternative sources of reliable evidence, such as observational or retrospective studies, have been lacking. The widespread adoption of electronic medical records (EMRs) and the increasing availability of computational methods to process vast amounts of unstructured information now make it possible to learn directly from practice-based evidence. We propose that “big data” clinical repositories, including radiology reports, can lend themselves to a treasure trove of point-of-care, relevant, actionable data that can be used in an innovative and cost-sensitive approach to evaluate the appropriate use of medical imaging. We aim to create a predictive model that leverages real-time EMR clinical data from top national medical centers to arrive at a patient-specific imaging outcome prediction. We recognize that clinicians have to make on-the-spot medical imaging-ordering decisions and they generally do not comply with existing clinical decision support rules. Our study aims to provide clinicians with a tool that can leverage aggregate patient data for medical imaging decision making at the point of care. The overarching approach of this study is to utilize scalable methodology that can be widely applied to leverage EMR data to predict the outcome of a several other high-cost, low-yield imaging tests. This proposal has the potential to better inform advanced imaging in the learning healthcare system of the future and reduce unnecessary imaging examinations and healthcare costs.