Early Recognition of Undiagnosed Viral Syndromes and Their Predictors Over Time

We propose to address the area of Emerging Viral Diseases by using computational methods to accelerate early identification of unusual viral syndromes. Overarching Challenge: The COVID-19 pandemic has exposed a critical need for more rapid identification of emerging infectious syndromes and characterization of epidemiological linkage. Earlier recognition of COVID-19 could have facilitated early containment and eradication efforts. Background: Since many existing methods for early identification of emerging viral disease require pre-existing case definitions and a volume of cases showing significant increase, better systems would leverage the rich, detailed information in the electronic health record (EHR). These instances are often documented by the astute clinician who suspects that not only is a case anomalous, but its presentation may also be consistent with emerging disease. In our work on COVID-19, we note that clinicians record distinctive features (e.g., travel from Hubei) and findings that may involve escalation (e.g., with public health authorities). While more such examples exist, there remain opportunities to surface the anomalous documentation from these clinicians and cluster their observations for earlier surveillance purposes. Research Plan: Leveraging data from the Department of Veterans Affairs (VA) EHR, which contains the records of over 20 million Veterans, we will apply methods drawn from natural language processing and machine learning to find features of unusual viral syndromes, cluster them, and characterize their evolution over time. Aim 1 of this proposal will identify distinctive features within EHR data such as exposures, risk factors, and expression of concern for possible emerging viral disease. Methods in this aim will include information retrieval, topic modeling, knowledge graphs, and a methodology we have developed for machine-assisted qualitative data analysis. Aim 2 will employ clustering methods to link anomalous case features and involve training machine learning classifiers such that disease can be identified within small case numbers. Aim 3 will utilize dynamic topic modeling and the QUICk ((QUalitative Interdisciplinary Collaboration) process to characterize concepts of viral disease over time to surface more timely and detailed information about the disease than what can be found in publicly available reports. Impact: The proposed approach will allow more rapid detection of emerging diseases while also being more detailed with improved characterization. This will permit earlier containment, earlier eradication, and improved care in the face of emerging viral disease. Military Relevance: As evidenced by COVID-19, emerging viral diseases represent a major threat to U.S. Veterans and Service Members, the nation, and the world. These viruses compromise their health, economic stability, well-being, and in the case of Service Members, their combat readiness. Our proposal seeks to address the gaps of current science in surveillance and our understanding of emerging respiratory pathogens by using careful application of machine learning. It may benefit Service Members and Veterans in the future by influencing early recognition, response, and improved outcomes. Less