Long COVID continues to challenge researchers as they grapple to understand how and why it happens. This is particularly perplexing in people who develop new and sometimes disabling symptoms after recovering from mild COVID-19, or who were even asymptomatic. Long COVID’s high variability from person to person, along with the millions of people now living with it—knowingly or unknowingly–make it extremely difficult to trace common factors of vulnerability.

On the National Institutes of Health (NIH) Director’s Blog, Lawrence Tabak, D.D.S., Ph.D. points to a new NIH-funded study that may help researchers find answers.*

Artificial intelligence (AI) may help

In Lancet Digital Health, a team of researchers led by Emily R. Pfaff, Ph.D., M.S. of the University of North Carolina Chapel Hill School of Medicine, and Melissa A. Haendel, Ph.D. of the University of Colorado Anschutz Medical Campus, Aurora, made a groundbreaking discovery. 

They found that a computer algorithm could reliably scan tens of thousands of electronic health records (EHRs) and identify those with Long COVID. While still preliminary, the research suggests that machine learning could help determine whether someone with COVID-19 is at risk for Long COVID.

Machine learning involves finding patterns by sorting through vast amounts of data. Machine learning can identify fine data patterns that people might overlook, because it doesn’t require humans to tell it what traits to search for.

What the researchers did

The researchers created an algorithm to search de-identified EHRs from COVID-19 patients included in the NIH”s National COVID Cohort Collaborative (N3C), a part of NIH’s RECOVER initiative (Researching COVID to Enhance Recovery) for better understanding Long COVID.

• Nearly 1.8 million living adult patients 18 years of age and older who had a COVID-19 diagnosis or positive test result at least 3 months prior were identified.
• In a sample of nearly 100,000 patients, common features including demographics, healthcare utilization, diagnoses, and medications were examined.
• The EHRs were inputted into the computer, as was health data from close to 600 Long COVID patients who had been treated at a post-COVID clinic.

Several machine learning models were created:

• One to find potential Long COVID among all COVID-19 patients
• A second to find potential Long COVID among all hospitalized COVID-19 patients
• A third to find potential Long COVID among all non-hospitalized COVID-19 patients

What they found

The algorithm proved highly accurate (85%) at finding patients with potential Long COVID. Important features from EHRs included healthcare use, age, and breathlessness (dyspnea), among others.

What it means

Tabak explains that once people with Long COVID can be identified from a large patient registry, researchers can start looking at differences in risk factors or treatments that might reveal why Long COVID happens in one person and not the next.

The study also highlights N3C’s usefulness for more effective study of COVID-19 and its long-term effects. It makes strides toward meeting RECOVER’s “urgent goal” of identifying people with Long COVID, or at risk for the condition, who may be eligible for new treatment studies.  

Tabak encourages people to enroll in RECOVER “to help researchers solve this puzzle.”

*Tabak, L. (2022, June 7). Using AI to Advance Understanding of Long COVID Syndrome. NIH Director’s Blog. https://directorsblog.nih.gov/2022/06/07/using-artificial-intelligence-to-advance-understanding-of-long-covid-syndrome