Nowcast Deep Learning Models For Constraining Zero-Day Pathogen Attacks – Application on Chest Radiographs to Covid-19

Abstract Outbreaks due to emergent pathogens like Covid-19 are difficult to contain as the time to gather sufficient information to develop a detection system is outpaced by the speed of transmission. Here we develop a general pneumonia (PNA) CXR Deep Learning (DL) model (MAIL1.0) follow by a second-generation DL model (MAIL2.0) for detection of Covid-19 on chest radiographs (CXR). We validate the models on two prospective cohorts of high-risks patients screened for Covid-19 reverse transcriptase-polymerase chain reaction (RT-PCR). MAIL1.0 has an Area Under the Receiver Operating Characteristics (AUC) of 0.93, sensitivity and specificity of 90.5% and 76.7% in detection of visible pneumonia and MAIL2.0 has an AUC of 0.81, sensitivity and specificity of 84.7% and 71.6%, significantly outperforming radiologists, especially amongst asymptomatic and patients presenting with early symptoms. Nowcast DL models may be an effective tool in helping to constrain the outbreak, particularly in resource-stretched healthcare systems.