Long COVID Diagnostic Score Under Fire: Researchers Warn Validation Study Overstates Accuracy

When a new diagnostic test emerges for a baffling disease, clinicians understandably want to know one thing: can we trust it? For Long COVID, a condition that has defied easy definition since the pandemic began, that question has taken on urgent importance. A scoring system known as the PASC score, developed through the National Institutes of Health’s RECOVER Initiative, promises to bring statistical rigor to a diagnosis that has until now rested largely on clinical judgment. But a new letter published in the Journal of General Internal Medicine argues that the score’s apparent accuracy has been substantially overstated, not because the underlying index is necessarily flawed, but because the study that put it to the test was built on a foundation of methodological bias.

The letter, written by Jason D. Goldman of the University of Washington’s Division of Allergy and Infectious Diseases and Jeffrey N. Martin of the Department of Epidemiology and Biostatistics at the University of California, San Francisco, takes aim at a validation study by Azola and colleagues. That study sought to determine how well the PASC score, formally known as the RECOVER-Adult Long COVID Research Index, performs when applied to real patients. The score assigns point values to various symptoms and laboratory findings, with a threshold above which a patient is classified as having Long COVID, also called post-acute sequelae of SARS-CoV-2 infection, or PASC. Goldman and Martin do not dismiss the effort. In fact, they commend the Azola team’s willingness to consider any symptom as a potential manifestation of the disease, in line with the broad definition advanced by the National Academies of Sciences, Engineering, and Medicine. Their objection concerns how the study’s accuracy estimates were produced.

At the heart of the critique lies a problem familiar to anyone who studies diagnostic testing: spectrum bias. When researchers evaluate how sensitive a test is, meaning its ability to correctly identify people who truly have the condition, the ideal case group should be a representative sample of all people with that condition in the general population. Some will have florid, multi-symptom disease; others will have a single lingering complaint. But the Azola study drew its Long COVID cases from a specialty Long COVID clinic, a population that Goldman and Martin describe as almost certainly enriched for patients with many symptoms. Because a positive PASC score requires a sufficient accumulation of symptom points, patients with only one symptom, or with several symptoms that carry low point values, would be far less likely to clear the diagnostic threshold. Had such patients been included in the case group, the measured sensitivity would have been considerably lower. By selecting cases from a clinic that preferentially treats the most symptomatic patients, the study effectively guaranteed that the score would look better at detection than it would in the wider world.

The bias did not stop there. Goldman and Martin note that the Azola study required its Long COVID cases to have at least one neuropsychiatric symptom, and that 62 percent of the case group experienced brain fog, a symptom worth three points under the scoring system. This requirement means the researchers built part of the case definition directly into the enrollment criteria. When a portion of what the test is designed to detect is already guaranteed to be present in every case, the test’s measured sensitivity is inflated by construction. It is a subtle but consequential distinction: the study did not ask whether the PASC score can find Long COVID among all people who have it, but whether it can confirm the disease in patients pre-selected to display some of its hallmark features.

Specificity, the flip side of diagnostic accuracy, fares no better in the analysis by Goldman and Martin. A test’s specificity reflects how reliably it returns a negative result in people who do not have the condition. To measure this meaningfully for a Long COVID classifier, the authors of the letter argue, the optimal comparison group would consist of patients who present with symptoms that overlap with Long COVID but whose illness stems from something other than prior SARS-CoV-2 infection. That is precisely the challenge clinicians face in practice: a patient arrives with fatigue, cognitive difficulties, or exercise intolerance, and the doctor must determine whether the cause is Long COVID or some other condition. A classifier is only useful if it can separate these possibilities. Instead, the Azola study assembled a group described as fully recovered and healthy controls, people with no symptoms at all. It surprised no one, least of all Goldman and Martin, that not a single control participant met the PASC score threshold. The resulting specificity estimate, effectively perfect on paper, is inflated in a way that tells clinicians almost nothing about the scenario they actually confront.

The third pillar of the critique concerns predictive values, numbers that often matter most to patients and physicians. Positive predictive value indicates the probability that a person with a positive test truly has the disease; negative predictive value indicates the probability that a person with a negative test truly does not. Unlike sensitivity and specificity, predictive values depend intimately on how common the disease is in the population being tested. Case-control designs, in which researchers hand-pick a group with the disease and a group without it, cannot estimate predictive values because the chosen proportions bear no necessary relationship to real-world prevalence. Goldman and Martin point out that the Azola validation, as a case-control study, simply cannot deliver these numbers. Whatever predictive figures might be quoted would be artifacts of how many cases and controls the investigators chose to enroll, not properties of the score in any actual clinical population.

Goldman and Martin do acknowledge one strength of the original work. The Azola team examined different combinations of fewer symptoms, reported in their supplemental materials, in an effort to explore how alternative configurations of the index might perform. This flexibility matters, because a rigid threshold may behave very differently across the heterogeneous presentations of post-viral illness. But the letter writers emphasize that every symptom combination evaluated within the same case-control framework inherits the same distortions. Optimizing symptom sets within a biased design does not remove the bias; it merely produces precision around an inaccurate estimate.

The implications extend well beyond a single academic dispute. Long COVID affects millions of people worldwide, and the absence of an objective diagnostic test has complicated everything from clinical care to disability determinations to the design of treatment trials. The RECOVER index was intended, in part, to standardize research classification so that studies across institutions could enroll comparable patient groups. If validation studies overstate the score’s accuracy, researchers may adopt the threshold with unwarranted confidence, misclassifying patients in ways that blur the very scientific questions the index was built to sharpen. Conversely, an honest accounting of the score’s limitations does not render it useless; it clarifies that the tool was designed and evaluated for research classification, not as a bedside diagnostic, and that its performance in general clinical populations remains an open empirical question.

The letter distills the matter into a stark summary. What the Azola study actually measured, Goldman and Martin conclude, is the fraction of patients cared for in a specialty Long COVID clinic, all of whom had at least one neuropsychiatric symptom, who met the PASC score threshold for the condition. That is a legitimate and informative measurement, but it is not validation of the score. True validation would require sampling cases from the full spectrum of disease severity in the general population and comparing them against symptomatic individuals whose ailments arise from causes other than SARS-CoV-2 infection. Only such a design could yield unbiased estimates of sensitivity, specificity, and predictive value.

For the field of post-viral illness research, the episode serves as a methodological cautionary tale at a moment when the pressure to produce actionable diagnostics is immense. Patients want answers, clinicians want tools, and health systems want standardized criteria. But the discipline of clinical epidemiology, as articulated in the textbooks Goldman and Martin invoke, exists precisely to insist that a test’s advertised accuracy reflects the populations and circumstances in which it will actually be used. Until the PASC score faces evaluation in a truly representative case group and against genuinely symptomatic comparator groups, its published performance figures should be read as optimistic upper bounds rather than established facts. The difference between those two readings could shape how Long COVID is diagnosed, studied, and treated for years to come.

Subject of Research: Validation methods and diagnostic accuracy of the RECOVER PASC score for Long COVID classification

Article Title: RE: Putting the PASC Score to the Test: Clinical vs. Statistical Accuracy in Long COVID Diagnosis (Azola et al. JGIM)

Article References: Goldman, J. D., & Martin, J. N. (2026). RE: Putting the PASC Score to the Test: Clinical vs. Statistical Accuracy in Long COVID Diagnosis (Azola et al. JGIM). Journal of General Internal Medicine. https://doi.org/10.1007/s11606-026-10852-2

Image Credits: AI Generated

DOI: 10.1007/s11606-026-10852-2

Keywords: Long COVID, PASC score, RECOVER Initiative, diagnostic accuracy, spectrum bias, sensitivity, specificity, predictive value, Journal of General Internal Medicine, SARS-CoV-2, clinical epidemiology, validation study