How physicians are using machine learning to predict COVID-19 mortality

Back

Advancements in artificial intelligence, automation, and machine learning are changing the landscape of healthcare as we know it. That means bridging the gap between technological innovation and medicine has never been more important.

A paper recently published in the International Journal of Medical Informatics by Furqan B. Irfan, Assistant Professor and Director of Research Development at MSU College of Osteopathic Medicine, and several of his colleagues highlights just how impactful physicians who work on both sides of the line can be.

Irfan frequently partners on research with Fahad Shabbir Ahmed, a physician-resident in pathology at Wayne State University. They share an interest in machine learning, translational science and clinical outcomes—a rare combination that led to the innovative methods used in this publication. Together with several team members from Aga Khan University Hospital and Medical College in Pakistan, Irfan and Ahmed created a novel deep neural network that predicts mortality for COVID-19 patients based on data from the first 24 hours of hospital admission.

The paper itself came about early on in the COVID-19 pandemic. Irfan and his colleagues wanted to do research on this patient population, but at the time, the data they were looking for wasn’t available yet.

It was only when they reached out to Maleeha Naseem, a Senior Instructor in the Department of Community Health Sciences at Aga Khan University Medical College, where both Irfan and Ahmed studied, that a plan started to form. “This is a great example of a project that was an international collaboration between Michigan and Pakistan,” Irfan said.

The research included information for 1,228 COVID-19 patients from Aga Khan University Hospital, which is a relatively small amount of patient data for creating a machine learning algorithm. Since it wasn’t possible to expand the data set vertically, the authors came up with a way to expand it horizontally. This involved a three-stage process, which ultimately became the Neo-V framework presented in the paper.

The first stage was using statistical analysis to see which variables were associated with COVID-19 mortality. From there, the team looked at all the significant variables and determined which ones could be combined into “neo-variables.” Once the neo-variables were determined, they were then fed into the deep neural network model to create the machine learning algorithm. The use of the deep neural network model, which is more complex than traditional machine learning models, is what really sets the Neo-V framework apart. “What this approach did was allow us to create a great algorithm in terms of the sensitivity and specificity,” Irfan said, especially when compared to conventional machine learning models.

The team also created a web app that gives hospital systems and physicians access to the algorithm so they can better plan care for COVID-19 patients. Though the team is collecting more data now that includes vaccination status and various treatments for COVID-19, the algorithm was created early on in the pandemic. This is part of what makes the algorithm so useful, Irfan said. “There are still a lot of people who are unvaccinated, especially the ones ending up in the ICU. So this algorithm really applies even today, and will continue to do so because in developing countries, there are still not enough vaccines to go around.” Even when there are enough vaccines, there will be people who choose not to have the vaccine, just like in the US.

Physicians can enter information for a patient gathered from their first 24 hours of admission, and get an objective prediction of mortality. From there, they can be more strategic about treatment, understand implications for ICU capacity, and prepare the patient’s family for the most likely outcome.  

The publication has led to collaborations within MSU as well as outside the university. Irfan and Ahmed now participate in University of Michigan’s Data Science Challenge group, and have also presented their Neo-V work to a machine learning forum at NIH National Center for Advancing Translational Sciences (NCATS).

“A lot of our work has to do with COVID-19 because that's of immediate importance right now. We've focused a lot of our energy and resources to help contribute,” Irfan explained. “But we've also made some breakthrough science work on cancer diagnostics and prognosis.”

For instance, machine learning may soon be able to predict which cancer patients will benefit from immunotherapy, and which ones should be spared from these therapies’ side effects because the results are unlikely to be positive. Beyond that, Irfan sees applications for machine learning algorithms in making key treatment decisions for all types of conditions, from COVID-19 and cancer to diabetes and high blood pressure, as well as more obscure conditions that are difficult to treat, like migraines. Plus, understanding more concretely which treatments are likely to work for each individual patient can save healthcare systems millions—maybe billions—of dollars, Irfan pointed out.

“This is the future of medicine,” he said. “We are happy to be at the forefront of creating such models, because they’re increasingly going to be used. It’s like being able to predict the future for the patient.”

Category:
Group: