Population-level Pulmonary Embolism Outcome Prediction with Imaging and Clinical Data: A Multi-Center Study

点击翻译按钮获取中文摘要

Project Summary Pulmonary embolism (PE) is a leading cause of death in the United States. Risk stratification for acute PE treatment can reduce mortality. Risk scoring systems use clinical and laboratory electronic medical record (EMR) data. In addition, biomarkers on computed tomography imaging can identify which patients with PE are at high risk of death, independent of clinical data. Despite advances in clinical and image-driven scoring systems, improving outcomes in acute PE depends on implementation of patient-specific EMR and imaging data analytic prognostic models at the point of care. The promise of digital medicine stems in part from the hope that by digitizing health data, we can leverage computer information systems to understand and improve care. A method that can make use of these data to predict patient-specific outcomes could not only provide major benefits for patient safety and healthcare quality but also reduce healthcare costs. Unfortunately, most of this information is not yet included in predictive statistical models that clinicians use to improve care delivery. This is because traditional computational methods and techniques are insufficient at accurately analyzing such high volumes of heterogeneous data. The goal of this proposal is to develop an automated precision medicine approach to achieve point-of-care risk stratification for PE patient outcomes using a fusion deep learning strategy that can simultaneously analyze health records and imaging data. An ideal PE risk-scoring system would not only predict mortality, but also assess the risk for the many debilitating long-term consequences of acute PE. Such a system would, therefore, facilitate optimal management and would likely require intelligent use of clinical, laboratory, and imaging data together in order to provide accurate patient -specific risk scoring for multiple PE outcome measures. In order to build a robust model, we propose to apply distributed training of deep learning models across four large US healthcare institutions. By distributing the algorithm rather than the data, we avoid sharing individually identifiable patient information. If successful, this project will be the first endeavor to leverage diagnostic imaging (pixel) data in combination with structured and unstructured EMR data to predict outcomes. We have the ideal research team, experience, and methods to develop an automated risk-scoring system for acute PE patients. Using a powerful combination of clinical, laboratory, and imaging data, this system will provide patient-specific risk scoring for multiple PE outcome measures. Further, this project will foster multi- center collaborations, which will afford us the opportunity to investigate the generalizability of our approach to different populations of PE patients and to train, test, and ultimately deploy our automated predictive model in a variety of clinical environments.