A promising new study published in the preprint online journal medRxiv in April 2020 shows the potential of artificial intelligence (AI) for developing a patient classifier that can separate patients likely to be negative for COVID-19 from among a pool of suspected patients visiting an emergency room (ER).
This would reduce the rate of spread significantly, by making it possible to immediately separate the patients most likely to be positive from others with similar symptoms of respiratory illness. It would protect both patients and healthcare providers from severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection.
The need for quick triage
The novel coronavirus (SARS-CoV-2) has spread across the world at unprecedented speed, placing a heavy and, in some cases, practically unsustainable, load on healthcare systems. Despite government aid, many healthcare providers find themselves requiring many more beds, intensive care units (ICU), and Personal Protective Equipment (PPE) than can be provided.
Efficient testing and diagnosis is an evident and essential requirement, but testing capacity is limited by the number of personnel, kits, and the time available. Many proposed solutions are being developed (including some that promote early detection through wearable technology), but the need has still not been completely met.
A quick and accurate diagnosis for those who visit the emergency room could improve the care of COVID-19 patients and help keep hospitals informed and prepared.
Viral detection tests
There are two main categories of tests: the first type detects the presence of a virus or protein, and the second type detects the antibodies manufactured by the body's immune system in response to the virus.
The first type, called molecular testing, uses a method called polymerase chain reaction testing or PCR. This is accurate, but labor-intensive, and therefore often cannot keep up to the pace required in the current coronavirus pandemic.
The second type of test is faster and less laborious, but is still under production, which means that it will take much time to scale up its numbers to the level required at present.
In the light of this, a team of scientists has come up with a low-cost and effective method of preliminary classification of suspected positives in a hospital ER based on simple and commonly done blood tests, to triage those who are most likely to be negative for the virus. The procedure is based on the analysis of the blood test results using artificial intelligence (AI).
How does the classifier work?
The working principle is relatively simple: the AI (called ER-CoV) accepts the results of widely available blood tests, analyzes them, and calculates the probability that the patient is infected with COVID-19.
Based on this probability, it recommends whether further, highly-sensitive testing is required. This could enable any lab with basic blood testing supplies to test people for coronavirus and recommend them for more sensitive testing.
One of the hurdles they had to overcome was that ER-CoV sometimes confused the input data for diagnoses of other respiratory diseases – especially influenza. To combat this, the team included a test for influenza in their proposed testing pipeline.
The team trained the Intelligence on public data collected from over five thousand patients at Albert Einstein Hospital, Brazil. They picked 599 samples with the greatest number of commonly performed tests such as leukocytes, mean platelet volume, mean corpuscular volume (MCV), creatinine, red blood cells, potassium, mean corpuscular hemoglobin concentration (MCHC) and red blood cell distribution width (RDW).
These patients were tested for COVID-19 using RT-PCR (reverse transcription-polymerase chain reaction), a testing method based on the detection of the genetic material of the SARS-CoV2 virus. 81 patients tested positive.
Based on this data, the AI compares the blood test results of ER cases to identify patients who have a high possibility of testing negative for the coronavirus and those who are "suspect cases."
How did the study perform?
The team reports that their system has an average specificity of 92.16% and a negative predictive value of 95.29%. They say that these results are "completely aligned with our goal of providing an effective, low-cost system to triage suspected patients at ERs."
Error testing showed that almost half of the 4% of cases where the AI reported a false negative result would be hospitalized anyway, meaning that the mistakes the AI makes include severe cases that would not be overlooked. This somewhat mitigates the test's lack of sensitivity.
Ability to limit testing and prevent spread
The team envisions an essential role for this simple, robust, and rapid screening technology in patient triage in the current world situation. Perhaps the most significant outcome of their system is in testing – at this point, many countries simply don't have the resources to test as many people as need it. The team predicts that they can reduce the number of tests done in emergency rooms by about 90%, with less than a 5% chance of getting a false negative.
The second important factor of the developed framework is concerning isolation. Currently, most hospitals keep people suspected of having COVID-19 in the same ward until test results are out, thereby increasing the chances of the virus spreading. The current testing model uses much less time than conventional methods, allowing the rapid screening of patients who have a high chance of carrying the virus and their isolation from those who don't until sensitive testing can be done for confirmation.
Moreover, the patients identified as positive for the virus by ER-CoV can then receive expedited testing, increasing the efficiency of their triage and, ultimately, their recovery.
medRxiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.
A novel high specificity COVID-19 screening method based on simple blood exams and artificial intelligence Felipe Soares, Aline Villavicencio, Michel Jose Anzanello, Flavio Sanson Fogliatto, Marco Idiart, Mark Stevenson medRxiv 2020.04.10.20061036; doi: https://doi.org/10.1101/2020.04.10.20061036