Understanding pediatric long COVID using a tree-based scan statistic approach: An EHR-based cohort study from the RECOVER Program

Vitaly Lorman, Suchitra Rao, Ravi Jhaveri, Abigail Case, Asunción Mejías, Nathan M. Pajor, Payal Patel, Deepika Thacker, Seuli Bose‐Brill, Jason P. Block, Patrick C. Hanley, Priya Prahalad, Yong Chen, Christopher B. Forrest, L. Charles Bailey, Grace M. Lee, Hanieh RazzaghimedRxiv 2022 · DOI

TL;DR

Researchers used a data-mining technique called a tree-based scan statistic to analyze Electronic Health Records (EHR) from over 14,000 children, identifying a broad spectrum of conditions associated with pediatric Long COVID (PASC). By letting the data "speak for itself" rather than relying on pre-defined clinical checklists, the study confirmed known symptoms like fatigue and shortness of breath while uncovering less-documented signals like hematuria and nutritional deficiencies.

What Problem Does This Paper Address?

Post-acute sequelae of SARS-CoV-2 infection (PASC), commonly known as Long COVID, is notoriously difficult to define in children. Unlike adults, pediatric cases often present with a confusingly wide range of symptoms and varying degrees of severity.

Most previous studies have been limited by small sample sizes or "clinician bias"—where researchers only look for symptoms they already expect to find. This creates a circular logic that might miss novel or under-reported manifestations of the disease. The authors of this paper sought to break this cycle by using an agnostic, data-driven approach to map the clinical landscape of pediatric PASC across a massive, multi-site national cohort.

Methods

The study employed a propensity-matched cohort design using data from the RECOVER Program. They compared three groups:

  1. PASC Cases (N=1,309): Children with specific ICD-10 diagnosis codes for PASC (U09.9 or B94.8).
  2. COVID-Positive Controls (N=6,545): Children who had COVID-19 but no PASC diagnosis.
  3. COVID-Negative Controls (N=6,545): Children who never tested positive for SARS-CoV-2.

The core innovation is the use of the tree-based scan statistic (TreeScan). Imagine the thousands of ICD-10 diagnosis codes arranged like a tree, where the trunk is "all diseases" and the leaves are highly specific conditions (e.g., "Migraine without aura, intractable"). The TreeScan algorithm "scans" every branch and leaf of this hierarchy simultaneously. It uses a likelihood ratio statistic to identify clusters where the observed number of cases in the PASC group significantly exceeds what would be expected based on the control groups, automatically adjusting for the "multiple testing" problem inherent in looking at thousands of variables.

Datasets

The researchers utilized the RECOVER PEDSnet EHR database, which aggregates data from nine major US children’s hospitals (including CHOP, Cincinnati Children’s, and Seattle Children’s). The dataset included clinical histories through December 2022, capturing a diverse range of demographics and clinical settings (inpatient, ED, and outpatient). To ensure they were looking at new conditions, they implemented an 18-month "washout period" prior to the initial infection.

Key Insights and Ideas

  • PASC is Truly Multi-Systemic: The study confirmed significant "clumping" of diagnoses across the cardiac, respiratory, neurologic, psychological, endocrine, GI, and musculoskeletal systems.
  • The "Symptoms" Cluster is the Strongest Signal: The most significant differences between PASC and controls were found in the "R00-R99" ICD-10 category (Symptoms and signs not elsewhere classified). This highlights that PASC often presents as a constellation of functional issues (like malaise or breathing difficulty) rather than a single, easily categorized organ disease.
  • Surprising "New" Signals: The data-driven approach flagged conditions that aren't usually the focus of Long COVID headlines, such as hematuria (blood in urine), nutritional deficiencies, and obesity. It’s unclear if COVID-19 causes these directly or if they represent an exacerbation of underlying issues during recovery.
  • Variant Evolution: A sensitivity analysis suggested that the "pre-Omicron" era was more enriched for metabolic and immunologic signals, while the "Omicron" era showed more signals related to the urinary system and migraines.

Key Results

  • Circulatory and Respiratory Dominance: Within the symptoms branch, the most significant "cuts" were for dyspnea (shortness of breath), chest pain, and fatigue/malaise.
  • Statistical Enrichment: In the PASC vs. COVID-negative comparison, the algorithm identified 304 significant clusters (cuts) in the diagnostic tree.
  • Validation of Complexity: A majority of children in the PASC cohort (55.8%) had at least one pre-existing chronic condition, but the propensity matching ensured that the PASC signals were distinct from their baseline health status.

Limitations and Future Work

The authors acknowledge several hurdles:

  • Coding Lag: The specific PASC code (U09.9) was only implemented in October 2021. Earlier cases relied on proxy codes, which might introduce bias.
  • The "Null" Bias: There are likely many children with Long COVID who never received a formal diagnosis code. These children would end up in the "control" group, which actually makes the study's findings more conservative (biasing results toward the null).
  • Tertiary Care Bias: The data comes from large academic health systems, which may not perfectly represent children treated in smaller community practices or those with less access to specialized care.

Future Work: The authors suggest that these identified clusters should now be used to derive "clinical phenotypes"—essentially, rules-based definitions that can help doctors identify and treat different "flavors" of pediatric Long COVID more effectively.

Summary updated

Abstract

STRUCTURED ABSTRACT Objectives Post-acute sequalae of SARS-CoV-2 infection (PASC) is not well defined in pediatrics given its heterogeneity of presentation and severity in this population. The aim of this study is to use novel methods that rely on data mining approaches rather than clinical experience to detect conditions and symptoms associated with pediatric PASC. Materials and Methods We used a propensity-matched cohort design comparing children identified using the new PASC ICD10CM diagnosis code (U09.9) (N=1309) to children with (N=6545) and without (N=6545) SARS-CoV-2 infection. We used a tree-based scan statistic to identify potential condition clusters co-occurring more frequently in cases than controls. Results We found significant enrichment among children with PASC in cardiac, respiratory, neurologic, psychological, endocrine, gastrointestinal, and musculoskeletal systems, the most significant related to circulatory and respiratory such as dyspnea, difficulty breathing, and fatigue and malaise. Discussion Our study addresses methodological limitations of prior studies that rely on pre-specified clusters of potential PASC-associated diagnoses driven by clinician experience. Future studies are needed to identify patterns of diagnoses and their associations to derive clinical phenotypes. Conclusion We identified multiple conditions and body systems associated with pediatric PASC. Because we rely on a data-driven approach, several new or under-reported conditions and symptoms were detected that warrant further investigation. LAY SUMMARY Pediatric long COVID in children does not currently have a precise clinical definition, in part due to its widely varying presentation in kids. By comparing children diagnosed with long COVID to children who had COVID-19 but were not diagnosed with long COVID, this study identified several groups of symptoms and conditions that are associated with pediatric long COVID. These findings can be used towards developing a precise definition of long COVID in children for use in future studies.
BibTeX
@misc{lorman2022understanding,
  author = {Vitaly Lorman and Suchitra Rao and Ravi Jhaveri and Abigail Case and Asunción Mejías and Nathan M. Pajor and Payal Patel and Deepika Thacker and Seuli Bose‐Brill and Jason P. Block and Patrick C. Hanley and Priya Prahalad and Yong Chen and Christopher B. Forrest and L. Charles Bailey and Grace M. Lee and Hanieh Razzaghi},
  title = {{Understanding pediatric long COVID using a tree-based scan statistic approach: An EHR-based cohort study from the RECOVER Program}},
  year = {2022},
  doi = {10.1101/2022.12.08.22283158},
  howpublished = {medRxiv},
}
Figure