COVID-19 Rapid Diagnosis Decision Support Tool
This rapid diagnosis decision support tool uses common clinical symptoms and lab tests to provide the positive / negative predictive value of a patient having COVID-19 infection. Please review the important considerations below prior to use.
COVID-19 Model Details
This model was created using data from Surgisphere’s real time global research network comprised of more than 1,200 healthcare organizations. A total of 6,103 patients with PCR-confirmed COVID-19 infection were compared with a pool of almost 60,000 patients with other respiratory ailments (such as influenza, viral / bacterial pneumonia, the common cold, etc.). Initial clinical presentation (i.e. fever, sore throat, respiratory symptoms), lab tests (CBC with differential, LDH, troponin I, and comprehensive metabolic profile), and imaging studies (chest X-ray) were measured. See our Frequently Asked Questions below.
Machine Learning Algorithm
A machine learning decision tree model was created to determine the likelihood of a patient with specific findings to have COVID-19 infection. 99.27% of cases can be diagnosed on the basis of the patient’s initial presentation, CBC with differential, and chest X-ray findings when using this tool. Based off feedback from the WHO, we have modified this tool to make it more applicable to developing healthcare systems that may not have access to more advanced lab tests. There are important limitations with decision tree analysis, which are discussed below in the Frequently Asked Questions.
Accuracy and Validity
This model has a positive predictive value of 99.90%, negative predictive value of 99.95%, sensitivity of 99.56%, and specificity of 99.99% for an overall accuracy of 99.95%. This model was prospectively validated on 4,780 patients. 37 hospitals from nine countries provided data for this model. A clinical research paper that describes the Methods and Results has been submitted for publication. This model is made available here in the interest of helping patients. Additional details will be made available after the paper has been accepted for publication.
The use of this machine learning decision support tool can help identify patients with presumed COVID-19 infection while waiting for PCR results to come back. This allows timely delivery of healthcare, earlier isolation of patients to reduce transmission, more effective mobilization of healthcare resources, and a positive public health impact.
Frequently Asked Questions
How Can I Contribute?
Surgisphere is dedicated to making the world a better place – for all people. Our research collaboration has helped improve the quality of care for kidney failure patients, improve outcomes for patients with cardiovascular disease, and avoid unnecessary hospitalizations and clinical testing for thousands of people. This effort requires participation of many good people – hospitals that want to use data to drive better outcomes, physicians who want to make more informed clinical decisions, insurance providers who want to make healthcare more accessible, and patients who want better outcomes.
If you are a physician, consider contributing to this effort by joining the research collaboration. There is no charge to participate.
If you represent a hospital, consider joining our global healthcare collaborative. You’ll get high quality dashboards and insight into clinical care that helps you deliver higher quality care for a more affordable cost. There is no charge for you to participate.
We are preparing a tool that helps predict outcomes for patients infected by COVID-19. What is the risk of serious illness? What is the chance of death? These are difficult questions to answer, but ones that we think we can solve using big data and machine learning. Consider funding this effort with a small donation.
This is a machine learning model derived from 6,103 patients. While machine learning is a very powerful tool, deriving results from this limited sample size limits introduces bias and can greatly affect the results. Incomplete lab testing at some clinical sites can also bias the findings. There are also variations in standard laboratories between clinical sites. THIS MACHINE LEARNING BASED DECISION SUPPORT TOOL SHOULD NOT BE USED IN LIEU OF SOUND CLINICAL DECISION MAKING AND PROVEN TESTING METHODS. ALWAYS CONSULT WITH A LICENSED AND COMPETENT PHYSICIAN, PARTICULARLY AS IT RELATES TO THIS INTERNATIONAL PUBLIC HEALTH EMERGENCY.