Machine Learning Tool May Help ED Clinicians Rule Out COVID-19
Machine learning can accurately predict the probability of COVID-19, allowing ED clinicians to reduce the number of patients referred for testing.
Source: Getty Images
By Jessica Kent
- A machine learning algorithm can detect the likelihood of COVID-19 infection using routine blood tests, potentially lowering the number of patients referred for PCR testing in the ED, according to a study published in the Journal of Medical Internet Research.
PCR testing is the current standard diagnostic for COVID-19. The process requires specific sampling like a nasal swab, as well as specialized equipment to run. PCR tests have been in short supply in some locations during the pandemic, causing delays in results, researchers noted.
"According to data from over 100 US hospitals, the national average turnaround time for COVID-19 tests ordered in emergency rooms is above 24 hours, far from the targeted one-hour turnaround,” said Tanya Kanigan, PhD, Biocogniv chief operating officer.
While providers have developed more effective and efficient ways of treating patients with the virus, testing shortages and delayed results can hinder any progress the industry has made.
"Nine months into this pandemic, we now have a better understanding of how to care for patients with COVID-19, but there's still a big bottleneck in COVID-19 diagnosis with PCR testing," said lead author and University of Vermont Assistant Professor Timothy Plante, MD, MHS.
Complete Blood Count and Complete Metabolic Panels are common laboratory tests ordered by emergency departments. These tests have a rapid turnaround time, and provide insight into a patient’s immune system, electrolytes, kidney, and liver.
The team was able to train a machine learning model developed by Biocogniv called AI-CVOID. The algorithm analyzed changes in these routine tests and assigns a probability of the patient being negative for COVID-19 with high accuracy. Researchers trained the model using 2183 PCR-confirmed COVID-19 cases from 43 hospitals during the pandemic.
To validate the model, the group used real-world data from Cedars-Sinai as well as data from geographically and demographically diverse patient encounters from 22 US hospitals.
The results showed that the machine learning model achieved an area under the curve (AUC) of 0.91 out of 1.00.
"This enables the model to achieve a high sensitivity of 95 percent while maintaining moderate specificity of 49 percent, which is very similar to the performance of other commonly used rule-out tests," said Biocogniv Chief Scientific Officer George Hauser, MD, a pathologist.
The model has the potential to accelerate clinical care processes in the ED.
"AI-COVID takes seconds to generate its informative result once these blood tests return, which can then be incorporated by the laboratory into its own test interpretation," said Jennifer Joe, MD, an emergency physician in Boston, Massachusetts and Biocogniv's Chief Medical Officer. "In an efficient emergency department that prioritizes these routine blood tests, the door-to-result time could be under an hour."
Additionally, the researchers believe that laboratories that incorporate the AI-COVID model could reduce the time it takes to deliver traditional PCR results.
"With the help of AI-COVID, laboratories might relieve some of the testing bottleneck by helping providers better allocate rapid PCR testing for patients who really need it," said Joe.
Clinical decision support tools like these have consistently helped ED clinicians rule out possible diagnoses, the team noted.
"For example, a low D-dimer blood test can help us rule out clots in certain patients, allowing providers to skip expensive, often time-consuming diagnostics such as chest CT scans," said Cedars-Sinai pulmonary and internal medicine specialist Victor Tapson, MD.
As COVID-19 cases continue to rise across the country, artificial intelligence and machine learning tools will play a critical role in helping providers prioritize care and understand patients’ needs.
"I'm honored to have such an impressive team of medical scientists from the University of Vermont and Cedars-Sinai as collaborators in validating this timely model,” said Biocogniv CEO Artur Adib, PhD.
“AI has progressed considerably; the time is now to leverage this powerful tool for new healthcare breakthroughs, and we're glad to direct it to help hospital laboratories and providers combat the current COVID-19 crisis."
