Towards an Artificial Intelligence Framework for Data-Driven Prediction of Coronavirus Clinical Severity

摘要

The virus SARS-CoV2, which causes the Coronavirus disease COVID-19 has become a pandemic and has spread to every inhabited continent. Given the increasing caseload, there is an urgent need to augment clinical skills in order to identify from among the many mild cases the few that will progress to critical illness. We present a first step towards building an artificial intelligence (AI) framework, with predictive analytics (PA) capabilities applied to real patient data, to provide rapid clinical decision-making support. COVID-19 has presented a pressing need as a) clinicians are still developing clinical acumen given the disease's novelty, and b) resource limitations in a rapidly expanding pandemic require difficult decisions relating to resource allocation. The objectives of this research are: (1) to algorithmically identify the combinations of clinical characteristics of COVID-19 that predict outcomes, and (2) to develop a tool with AI capabilities that will predict patients at risk for more severe illness on initial presentation. The predictive models learn from historical data to help predict specifically who will develop acute respiratory distress syndrome (ARDS), a severe outcome in COVID-19. Our experimental results based on two hospitals in Wenzhou, Zhejang, China identify features most predictive of ARDS in COVID-19 initial presentation which would not have stood out to clinicians. A mild increase in elevated alanine aminotransferase (ALT) (a liver enzyme)), a presence of myalgias (body aches), and an increase in hemoglobin, in this order, are the clinical features, on presentation, that are the most predictive. Those two centers' COVID-19 case series symptoms on initial presentation can help predict severe outcomes. Predictive models that learned from historical data of patients from two Chinese hospitals achieved 70% to 80% accuracy in predicting severe cases.