A real-world intelligent system for the diagnosis and triage of COVID-19 in the emergency department

Abstract

The Coronavirus Disease 2019 (COVID-19) pandemic has had an unprecedented impact on healthcare systems, prompting the need to improve the triaging of patients in the Emergency Department (ED). This could be achieved by automatic analysis of chest X-rays (CXR) using Artificial Intelligence (AI). We conducted a research project to generate and thoroughly document the development process of an intelligent system for COVID-19 diagnosis. This work aims at explaining the problem formulation, data collection and pre-processing, use of base convolutional neural networks to approach our diagnostic problem, the process of network building and how our model was validated to reach the final diagnostic system. Using publicly available datasets and a locally obtained dataset with more than 100,000 potentially eligible CXR images, we developed an intelligent diagnostic system that achieves an average performance of 93% success. Then, we implemented a web-based interface that will allow its use in real-world medical practice, with an average response time of less than 1 second. There were some limitations in the application of the diagnostic system to our local dataset which precluded obtaining high diagnostic performance. Although not all these limitations are straightforward, the most relevant ones are discussed, along with potential solutions. Further research is warranted to overcome the limitations of state-of-the-art AI systems used for the imaging diagnosis of COVID-19 in the ED.