Vironix: A Machine-Learned Approach to Remote Screening, Surveillance, and Triage of Viral Respiratory Illness

Abstract

RATIONALE The Covid-19 pandemic has posed a serious, ongoing global health challenge. The United States has been the worst affected, with more than 11M confirmed cases and 246K deaths (as of November 2020). Two primary and persisting concerns are the continued necessity for shutdown/isolation and the possibility of singular waves of rapid virus spread that could overwhelm global healthcare systems, resulting in preventable mortality and substantial economic burden. While vaccines are being developed and disseminated, the need for remote patient care has never been more critical. To that end, we developed a Covid-19 remote triage software, Vironix, which uses machine-learning algorithms to enable real-time risk stratification and decision support for users. This remote management approach has significant potential to increase safety, improve health outcomes, and stem virus spread as organizations reopen. METHODS Vironix uses personalized machine-learning algorithms trained off clinical characteristic data from the EU, East Asia, and the USA in tandem with prescribed guidelines from the CDC, WHO, and Zhejiang University's handbook on Covid-19 prevention. Clinical characteristics of thousands of patients in the literature were mapped into patient vignettes using Bayesian inference. Subsequent stacked, ensemble decision tree classifiers were trained on these vignettes to classify severity of presenting symptoms and signs. Crucially, the algorithm continuously learns from ongoing use of the application, strengthening decisions, and adapting decision boundaries based on inputted information. Vironix was deployed using a user-friendly API, allowing users to easily screen themselves and obtain remote decision support through a variety of devices (mobile apps, computers, health monitors, etc).RESULTS Algorithm performance was assessed based on its binary classification performance in an out-of-sample test set including severe and nonsevere labels. Vironix correctly assigned the severity classes with an accuracy of 87.6%. Vironix further demonstrated superior specificity (87.8%) and sensitivity (85.5%) in identifying positive (severe) presentations of Covid-19. The algorithms, deployed behind the Vironix Web Application, have been invoked by tens of thousands of users around the world. CONCLUSION 1. The Vironix approach is a highly novel, generalizable methodology for mapping clinical characteristic data into patient scenarios for the purpose of training machine-learning prediction models to detect health deterioration due to viral illness. 2. Vironix exhibits excellent accuracy, sensitivity, and specificity in identifying and triaging clinical presentations of Covid-19 and the most appropriate level of medical urgency. 3. Algorithms continuously learn and improve decision boundaries as individual user input increases. .