The U.S. needs early warning systems to help it contain the spread of infectious diseases . Conventional early warning systems use lab-test results or dynamic records to signal early warning signs . New early warning systems can supplement these data with indicators of public awareness like news articles and search queries . This study aims to explore the potential of utilizing social media data to enhance early warning of the COVID-19 outbreak . To demonstrate the feasibility, this study conducts a retrospective analysis and investigates more than 14 million related Twitter postings in the date range from January 20 to March 10 , 2020 . With the aid of natural language processing tools and machine learning classifiers, this study classifies each of these tweets into either a signal or a non-signal . In this study, a 'signal' tweet implies that the user recognized the COVID-19 outbreak risk in the U.S . This study then proposes a parameter 'signal ratio' to signal warning signs of the COVID-19 pandemic over periods . Results reveal that social media data and the signal ratio can detect the hazards ahead of the COVID-19 outbreak . This claim has been validated with a leading time of 16 days through the comparison to other referenced methods based on Google trends or media news.