Loading ...

As healthcare systems continue to grapple with identifying and managing long COVID, artificial intelligence (AI) is showing promise as an important tool that could one day expand scientists’ understanding and even lead to new diagnostics for the condition. 

Three recent studies demonstrated how machine learning and AI algorithms can be leveraged to process vast amounts of complex clinical notes, hospital data, and patient data. All three highlighted how AI could potentially tackle different aspects of long COVID and advance how clinicians identify, track, predict, and treat it.

Because of the way the SARS-CoV-2 virus binds to human cells, long COVID complications can develop almost anywhere, experts said — from the brain to the heart to the gastrointestinal system — causing upwards of more than 200 symptoms. Many of these symptoms can also be caused by other diseases and conditions, making diagnosis and treatment challenging.

“That’s hard from a medical point of view, because that’s not typical of how we think of most illnesses that we deal with,” said Fahad Razak, MD, internist at St. Michael’s Hospital and Canada Research Chair in Healthcare Data and Analytics at the University of Toronto in Toronto, Ontario, Canada.

“Probably all of our data underestimates the real population-level burden of how many people are affected by [long COVID], and I think many people suffer in silence,” he said.

Three Studies, Three AI Applications

In a study published in Med, scientists at Mass General Brigham developed a type of AI tool called precision phenotyping to analyze millions of data points from the electronic health records of nearly 300,000 patients across 14 hospitals and 20 community health centers in Massachusetts. The technique identifies and tracks symptoms and conditions linked to COVID-19 to distinguish them from other illnesses.

Scientists said the tool was nearly 3% more accurate than current diagnostic methods for long COVID. It allows for very detailed and precise analysis of information that could help with the challenging task of diagnosis, ensuring patients receive appropriate care. It is an example of AI’s ability to synthesize, curate, and sift through enormous volumes of information.

The system only considers long COVID if the symptoms cannot be explained by anything else in the patient’s medical history. After exhausting all other possibilities, the tool flagged about 22.8% of cases as long COVID.

“It’s been really, really challenging to define and diagnose long COVID. One of the reasons is because its symptoms are very heterogeneous, overlapping with many things,” said co-author Hossein Estiri, PhD, head of AI research at the Center for AI and Biomedical Informatics of the Learning Healthcare System at Mass General Brigham and associate professor of medicine at Harvard Medical School, Boston.

“The more we can find innovative ways of using AI to address these complex, evolving phenotypes, the better,” he said.

The algorithm is already public and is packaged in a software tool that can be implemented in different institutions across the US and internationally, according to Estiri, adding that the center is looking for more institutions to participate. At this stage, it is helping to advance research designed to better understand the condition through larger sample sizes and flag potential patients to enroll in future studies or clinical trials.

“One of the most difficult things about long COVID is that it can affect almost any organ system, and in many ways, it’s a mimicker of many other illnesses that we have to deal with,” said Razak, who was also the scientific director of the Ontario COVID-19 Science Advisory Table and coauthored dozens of papers that shaped the policy, public health, and clinical response to the pandemic. “This is an interesting example where AI could do something much more efficiently than any of us, individually, clinically could do.”

Data Sharing for Local Decision-Making

At the University of Pennsylvania’s Perelman School of Medicine, Philadelphia, researchers used a machine learning technique called latent transfer learning to gain a clearer picture of the specific healthcare burdens of long COVID among pediatric patients in different hospitals. The technique, which boosts statistical precision by analyzing information across hospitals, tracked the electronic health records of 432,165 young patients from eight pediatric hospital systems.

According to a study published in Patterns, researchers found that many patients fell into one of four subgroups: those with mental health conditions like anxiety, depression, and Attention-deficit/hyperactivity disorder; those with atopic/allergic conditions, including asthma; those with noncomplex chronic conditions; and those with complex chronic conditions like multisystem disorders. The technique also identified the type of care patients required and the impact these patient groups had on hospitals.

Long COVID is less common in children than in adults. But it involves a unique and understudied patient population group that is growing, gaining weight, and developing their mental and cognitive understanding of the world, explained co-author of the study, Yong Chen, PhD, professor of biostatistics in the Department of Biostatistics, Epidemiology, and Informatics with the Perelman School of Medicine.

“The number one message we try to pass is the awareness and the complexity of pediatric long COVID,” said Chen. “It’s extremely complicated in terms of the subtype of long COVID for children and adolescents for the reason that they are highly entangled by their developmental age.” A COVID infection has a nontrivial impact on the digestive system, for example, so how would outcomes be measured without being confounded by a child’s natural growth? Chen explained.

The transfer learning approach is much more precise because it adaptively incorporates data from other hospitals while accounting for differences in patient populations, hospital staffing, and equipment. Chen and his colleagues advocate for hospitals and health systems to work together and share data to facilitate more personalized care and improve responses to future public health crises.

“You could start to triage your patients into better categories and different follow-up and management approaches,” said Razak. “The ability to identify that based on complex, many millions of data points of information would greatly enhance clinical care.”

Predicting Long COVID

In a much smaller and more limited study, researchers in Italy used three different machine learning approaches to predict with up to 94% accuracy which patients would eventually develop pulmonary long COVID. Clinical data collected early in the pandemic from patients with COVID-19 hospitalized across four different Italian hospitals were analyzed. The different AI methods used in the study illustrated effective strategies for predicting long COVID, the scientists said, even when the patient sample size was small.

Researchers said these approaches could help healthcare professionals identify which patients were more susceptible to developing long COVID and provide support to mitigate the condition’s long-term impact. It would allow doctors to identify high-risk patients early, allowing a more tailored approach to care and management strategies, and help healthcare providers allocate resources more efficiently.

Coming to a Hospital Near You?

Electronic health records make these kinds of insights not only possible but also easy to deploy and almost immediately useful.

“20 years ago, this information would have been largely unactionable, because most of the healthcare encounters would have been largely pen and paper, handwritten information that cannot be extracted out to use for analytics,” said Razak.

“What these papers are harnessing is the many billions of dollars of investments that have already occurred within the Canadian and the US healthcare systems to largely digitize the system,” he said.

Yet experts agree there are important ethical and feasibility constraints in medical AI that need to be addressed before AI can be used broadly in clinical settings.

“There’s governance, privacy, cost, and technical barriers that I would say are all solvable, but they would take resources and will to do it,” said Razak. He noted that even at St. Michael’s Hospital, where he works — “one of the most data-advanced hospitals in Canada” — there is only one algorithm implemented in its clinical practice.

Still, these findings are promising and intriguing from a research perspective, and these algorithms may one day help clinicians directly in decision-making on patient care.

“Could they be implemented in the health system? I think they could be implemented pretty quickly,” said Razak. “Would clinicians know what to do with them to change the way they’re managing clinical care? No, not yet.”