Artificial Intelligence Tool Could Help Triage COVID-19 Patients in the ED

By Marilynn Larkin

May 21, 2021

NEW YORK (Reuters Health) - A new artificial intelligence (AI) tool predicted with up to 80% accuracy which COVID-19 patients would deteriorate within 96 hours, researchers report.

A preliminary version of the tool was tested in the emergency department (ED) during the first wave of the pandemic and produced accurate predictions in real-time, showing potential to assist front-line physicians in triaging these patients.

"The multi-modal system uses the patient's clinical variables as well as chest X-rays, and the overall prediction is most accurate and representative of the patient's status when it is computed using the most recent data," Dr. Farah Shamout of NYU Abu Dhabi in the United Arab Emirates told Reuters Health by email.

"Therefore, the ED physician could prompt nurses to collect the patient's vital signs and blood tests, and radiologists to collect a new chest X-ray, to get an updated prediction," she said.

The deterioration prediction score is automatically computed, once the data are collected, she added. Further, she said, "It's easy and possible to fine-tune the model to predict deterioration among patients with other diseases."

As reported in npj Digital Medicine, the team trained their AI prognosis system on 5,224 chest X-rays taken from 2,943 seriously ill patients infected with SARS-Co-V-2. Clinical variables such as demographics, vital signs and lab results were also included; further, the need for mechanical ventilation and whether each patient went on to survive (2,405 patients) or die (538) were also factored into the mathematical models.

When predicting deterioration (e.g., need for mechanical ventilation) within 96 hours, the system achieved an area under the receiver operating characteristic curve of 0.786 - i.e., an accurate prediction for four out of five patients.

Real-life deployment of a preliminary version of the tool at NYU hospitals early in the pandemic suggested it could produce accurate predictions in real time.

Dr. Shamout noted, "We are currently interested in exploring prospective validation and external validation of our system to ensure that it performs well in clinical practice and generalizes across patient populations that were not seen during model development."

The model is publicly available for the research community at

Dr. Albert Hsiao, director of the Augmented imaging/Artificial intelligence Data Analytics Laboratory at the University of California, San Diego, told Reuters Health by email, "This work is interesting, and parallels some of our work implementing a similar system at UCSD Health. I agree with the authors that x-ray is important in these patients, and that AI algorithms could benefit how we triage patients with COVID-19."

"In our work, we also went on to assess how ED doctors felt about using our algorithm in their practice, to see if it had any benefit at all," he noted. "We were surprised that they found it useful in a quarter of their patients." (

"Future work should focus on how to discriminate between different entities, such as viral pneumonia and pulmonary edema, which can look quite similar on x-ray, because they can make an impact on the kinds of monitoring and treatments patients will need," Dr. Hsiao concluded.

SOURCE: Npj Digital Medicine, online May 12, 2021.