Researchers from Mount Sinai are the first in the United States to create and use an artificial intelligence (AI) algorithm to diagnose patients with COVID-19.
The algorithm can detect a COVID-19 case based on how lung disease looks in CT scans of a patient’s chest and data on the patient’s symptoms, age, bloodwork and possible contact with the virus.The team created the algorithm using CT scan data from coronavirus patients in Chinese medical centers. It was found to be just as accurate as experienced radiologists and sometimes better in cases with no clear sign of lung disease in CT scans.
Source: BioMedical Engineering and Imaging Institute (BMEII) at the Icahn School of Medicine at Mount Sinai
This research expands on a previous study that identified the characteristic pattern of the disease in the lungs of COVID patients and how it develops.
To develop the algorithm, the team used scans from 900 patients. These patients were admitted to 18 medical centers in 13 Chinese provinces between January 7 and March 3, 2020. According to the data, 419 of the patients were confirmed COVID-19 positive cases and 486 patients were COVID-19 negative. Researchers also had the patient’s clinical information including blood test results that showed abnormalities in white blood cell counts or lymphocyte counts, age, sex and symptoms. They focused on CT scans and blood tests because doctors in China used those tests to diagnose COVID-19 cases.
The algorithm mimics the workflow that a physician uses to diagnose COVID-19 to predict a positive or negative diagnosis. It produces separate probabilities of COVID-19 diagnosis based on CT images and clinical data. The algorithm was initially trained and fine-tuned on data from 626 out of 905 patients and tested on the remaining 279 patients that were split between positive and negative cases. The higher the sensitivity of the algorithm, the better detection performance it had.
The results showed that the algorithm had a significantly higher sensitivity of 84% compared to the 75% of typical radiologists’ evaluating data. The algorithm improved the detection of COVID positive patients who had negative CT scans. It recognized 68% of COVID positive cases where radiologists interpreted the cases as negative. Improved detection is important to slow the spread of the disease.
Imaging is not widely used for COVID-19 diagnosis in the U.S., but the researchers say it could play an important role in rapid and accurate diagnosis. The AI algorithm could provide a second opinion in cases with negative CT scans.
The team is now focused on further developing the model to determine how well COVID patients will fare based on clinical data and CT data to optimize treatment and improve their outcomes.
A paper on the AI algorithm was published in Nature Medicine.
