Building a Predictive Analytics Tool to Triage COVID-19 Patients

Of the many truths brought to light during the pandemic, the importance of providers’ voices is among the most consequential.

The experiences and daily operations of frontline healthcare workers can inform the development of helpful, user-friendly tools, both during and after the current crisis.

At the Feinstein Institutes for Medical Research – the science arm of Northwell Health – researchers were quick to tap into the clinician experience. The team kept providers in mind when developing a predictive analytics tool that could help triage COVID-19 patients. 

Using data from emergency department patients, including demographics, labs, and vitals, researchers designed an algorithm that can predict whether patients have a high risk of respiratory failure within 48 hours.

“Since the start of the pandemic, we've been in close contact with the frontline workers at Northwell who have been caring for COVID-19 patients in our emergency departments,” Theodoros Zanos, PhD, assistant professor at the Feinstein Institutes for Medical Research, told HealthITAnalytics.

“Very early on, we recognized that clinicians needed a way to identify which patients may deteriorate faster, as well as who would need ventilator support. We deliberately chose this window of 48 hours based on physician feedback.”

The tool is also designed to be easily integrated into clinical workflows.

“The tool can be embedded into each hospital’s EHR, and it can provide a patient’s risk of respiratory failure immediately to users. So, the tool can be part of the regular workflow that hospitals have within their ED or on the floor,” he said.

Respiratory failure is the leading cause of death among patients with COVID-19, Zanos said. It’s crucial for providers treating these patients to have the most accurate, up-to-date information available to deliver the best possible care.

“In general, predictive analytics are useful for physicians to augment their decision-making. When a physician sees something that makes her believe a patient might be deteriorating, that can help her make better decisions on how to care for them,” Zanos said.

“During the pandemic, there is an added urgency to these types of tools, because even now we keep on learning more and more about the virus and the trajectory of the patient from the moment that they show up in the ED all the way to discharge. Physicians have told us that many COVID-19 patients deteriorate very rapidly, so any signal that could prepare clinicians and give them time to make sure they have the necessary tools available is helpful.”

Zanos pointed out that the pandemic has also substantially increased clinicians’ acceptance and use of analytics solutions, as well as the development of these algorithms.

“One reason for this increase is the urgency of the situation and physicians’ need for such tools. We don’t need to convince providers that something like this could help them, because they actually come to us themselves and tell us that they need predictive solutions,” Zanos said.

“COVID-19 is also a good use case for these types of AI approaches, because there is a very definite diagnosis. And that helps a lot in building these models, because we can identify the specific cohorts that are needed to develop analytics tools. It's a combination of the need, the doctors being more familiar with these tools, and the nature of the problem that facilitates the development of these tools.”

The clinical tool designed by Zanos and his team achieved an average predictive accuracy of 92 percent, accurately identifying at-risk patients in need of earlier interventions. The group plans to implement the solution within some Northwell Health hospitals.

“The next steps are to try and embed it into the workflow of our hospitals and our emergency departments so that doctors can easily use it. We are also planning on developing an outward-facing website that would allow other physicians to go in and input some of the predictors of respiratory failure, and then they would get a patient’s probability outcome,” said Zanos.

“That website won't be as complete as the full model that we published, because the model incorporates a lot of different inputs. We don’t expect somebody to go on a website and start writing 15 or 20 different inputs – that will take too much time. But we can stratify and prioritize the inputs that are the most predictive, and we have done that with our model.”

A major contributor to the model’s success was the volume and diversity of the data used to train the tool – something that should be included in algorithm development going forward.

“The reason we believe the model is powerful is that it relies on one of the largest COVID-19 patient databases in the country. The dataset is also one of the most racially and socioeconomically diverse, because unfortunately we were right in the epicenter of the first wave of the pandemic in New York,” Zanos concluded.

“There’s a major benefit to that. These are things that make an AI algorithm a lot more robust and less biased towards a certain type of patient.”