By Paul Nicolaus
January 27, 2022 | Researchers at Mount Sinai and the University of Colorado are pursuing the possibilities that stem from predicting which COVID-19 patients are most likely to deteriorate over time. The hope is that this ability could help health care professionals more effectively triage patients and pair them with appropriate treatments to improve outcomes.
COVID Complexities and Patient Delineation
A group at Mount Sinai in New York has used a data mining technique called sequence cluster analysis to come up with four clinical subphenotypes in critically ill COVID-19 patients.
Their findings, detailed in a study published Nov. 23 in the Journal of the American Medical Informatics Association (DOI: 10.1093/jamia/ocab252), involve the analysis of over 1,000 critically ill patients with laboratory-confirmed SARS-CoV-2 infection. The researchers identified subgroups based upon biomarkers and treatments during the initial 24 hours of ICU stay, made possible using a technique called sequence cluster analysis.
The approach allows multiple features to be considered over a stretch of time, making it possible to separate subphenotypes even if some appeared clinically similar at admission or by the end of the observation period. Sequence clustering essentially deconvolutes the trajectory of patients in one-hour intervals, senior author Girish Nadkarni, professor of medicine at the Icahn School of Medicine at Mount Sinai, told Diagnostics World.
The sequence clustering algorithm usually comes from the realm of genomics and is seldom used in data mining and electronic health record research, added Nadkarni, who is also chief of the Division of Data-Driven and Digital Medicine, co-director of the Mount Sinai Clinical Intelligence Center, and clinical director of the Hasso Plattner Institute for Digital Health.
In this instance, though, the method allowed the research group to assess changes in patient status over one-hour intervals, see how patients evolved during that trajectory over time, and determine if any groupings could be made.
The logical future extension of this work would be to help match patients to specific treatments. If some patients are more likely to progress more quickly, more invasive or high-risk therapies might be considered earlier on. But at this stage, more work is needed, Nadkarni said, including randomized trials.
Moving forward, the plan is to pursue additional research involving longer timeframes and to take into account the follow-up after patients are discharged from the ICU. This would allow the researchers to not only predict and determine subphenotypes of patients within hospitals but also after they get discharged from the hospital, he said, which may allow for a better understanding of post-acute sequelae SARS-CoV-2 infection (PASC), also known as long COVID.
“We tend to think of diseases as labels and as single monoliths, but within every complex disease—and COVID is definitely a complex disease—there are multitudes,” said Nadkarni. Rather than thinking of COVID as a broad label, such as Type 2 diabetes or hypertension, it may be possible to get down into subgroups of a disease, each one of which may have a different outcome, and each one of which could be matched to a different treatment.
Gathering Insights for Diagnosis, Prognosis, and Triage
In another line of work, University of Colorado researchers have discovered genetic biomarkers that reveal not only who is infected with SARS-CoV-2 but also predict how severe COVID-19 disease will become, which they say could fill an important gap in diagnostic testing.
The study, published Oct. 26 in Communications Medicine (DOI: 10.1038/s43856-021-00042-y), suggests that signals from a process called DNA methylation differ between those who are infected with SARS-CoV-2 and those who are not. And these signals can indicate disease severity, even in the early stages.
The University of Colorado researchers and colleagues at Illumina customized the company’s Infinium MethylationEPIC array to analyze the epigenome using blood samples from people with and without COVID-19 (164 COVID-19 patients and 296 controls). Using statistical and machine learning techniques, they identified markers of SARS-CoV-2 infection in addition to the severity and progression of COVID-19 disease.
The cross-validated best fit AUC was 94% for case-vs-control status. For hospitalization, ICU admission, and progression to death, predictive findings were 79%, 81%, and 84%, respectively. The researchers concluded that the COVID-19-specific epigenetic signature offers insights that could be useful for diagnosis, prognosis, and triage.
This study is seen as “a first step toward selecting biomarkers for inclusion on a high-throughput methylation beadchip array specifically for the clinical diagnosis of COVID-19 disease that is also cost-effective given the added value of predicting subsequent clinical outcomes,” they noted in the study.
In the last couple of years, a lot has been learned about the power of leveraging epigenetics to generate diagnostic, prognostic, and predictive signatures, specifically in methylation, explained Kathleen C. Barnes, professor of medicine and epidemiology at the University of Colorado Anschutz Medical Campus.
“Before we launched this project, we were really inspired by work that others had done in partnership with Illumina and their commercially available methylation platform, the EPIC chip, specifically around differential diagnosis of neurotumors,” she told Diagnostics World.
She and colleagues have also been interested in generating epigenetic signatures around germline-based disease. So when the pandemic hit, Barnes reached out to Illumina and discussed the possibility of identifying epigenetic signatures in blood samples from patients infected with SARS-CoV-2. There was already literature on methylation changes with other infectious diseases, so it seemed like a reasonable endeavor to pursue.
Initially, “there was kind of a feeding frenzy,” Barnes explained, in terms of standing up COVID-19 diagnostic platforms. What is unique about this platform is that in addition to demonstrating accurate diagnosis of infection, she and colleagues were able to develop classifiers that predicted various outcomes associated with disease severity, such as those who go on to a ventilator and those who do not survive.
“If providers could know upfront that a patient presenting to them who has been newly diagnosed with COVID-19 was going to go on to develop worse disease, there are different steps that provider could take to triage that patient and ideally end up with a much more favorable outcome,” she said.
Scalability Considerations and Looking Beyond COVID
One of the challenges faced in imagining how this platform could be commercialized, marketed, and widely used as a competitive diagnostic test is that there was seemingly no way to reduce the turnaround time of several days.
However, in recent months, there has been a realization that monoclonal antibodies can reduce severity. And even more recently, antivirals are coming down the pike. Barnes and colleagues believe this provides a unique opportunity for providers to use a predictive algorithm to make educated decisions about who to treat.
Although many would-be patients present to the emergency department, some of those patients will be sick enough that they eventually wind up in the ICU. And a significant proportion of the population may develop infection and not present to a hospital initially because their symptoms are so mild. Even so, some of those people will go on to experience severe illness. “This could be a really robust tool for making a prediction about who is going to develop worse disease,” she said.
Another notable aspect of the work detailed in this study, according to Barnes, is the notion that this platform is not limited to SARS-CoV-2. It is also applicable to many other infectious diseases. As such, the researchers imagine it could be utilized with both new and emerging strains of SARS-CoV-2 as well as future pandemics.
“By building a database of individuals and these epigenetic signatures, we would be able to very quickly identify individuals who have a new viral strain when comparing those that we’ve already characterized,” she said. “So we believe it has broad significance in the context of not just COVID-19 but for future strains of COVID-19, for individuals who develop long haul disease, and other infectious diseases.”
The platform upon which this test is run is used in clinical laboratories across the United States, Barnes explained, and multiple samples can be tested at a time. It is not just about the commercial methylation chip used but also the ability to use machine learning to develop disease classifiers based on the patterns they see off of this chip.
“And to that end,” she said, “the sky is really the limit in terms of how many different disease classifiers we can develop.” It all depends on the extent to which they have biospecimens representing patients with different diseases and outcomes.
This study was conducted using peripheral blood samples, she explained, but the intent is to move forward with testing that enables the identification of the same signatures from DNA extracted from a saliva sample, which would be a far easier sample for a would-be patient to deliver. An additional scalability feature they are exploring is a home-based test that could be used to help reduce turnaround time.
When asked what’s next, Barnes pointed out that she and colleagues are continuing to partner with Illumina, the company that initially developed the methylation chip. And she now wears two hats, both as a faculty member at the University of Colorado (where the work took place) and as senior vice president at Tempus.
“In that space, we have tremendous opportunity to really scale the development and validation of these disease classifiers because Tempus is a company predicated on big data and leveraging AI to come up with predictive algorithms that diagnose and predict disease,” she said. “And so both through our partnership with the university, with Illumina, and at Tempus, we are moving forward with amassing additional biospecimens from patients.”
That will make it possible to begin to disentangle the diagnosis of SARS-CoV-2 infection and the prediction of worsening disease. It will also enable the characterization of disease classifiers in real-time as new strains emerge and make it possible to zero in on individuals with long haul syndrome. “Our goal right now is to scale all of this,” she said, while also exploring how this platform can be applied to other infectious diseases.
What has been learned over the past year or two is that epigenetics “has tremendous power” in not only diagnosing but also making predictions about disease that is not possible using traditional genomics alone, Barnes said. And it has broad application across different types of diseases. “This really feels like a frontier,” she added, “and we’re excited to be part of that.”
Paul Nicolaus is a freelance writer specializing in science, nature, and health. Learn more at www.nicolauswriting.com.