Pediatric Long COVID Subphenotypes: An EHR-based study from the RECOVER program

Abstract

Pediatric Long COVID has been associated with a wide variety of symptoms, conditions, and organ systems, but distinct clinical presentations, or subphenotypes, are still being elucidated. In this exploratory analysis, we identified a cohort of pediatric (age <21) patients with evidence of Long COVID and no pre-existing complex chronic conditions using electronic health record data from 38 institutions and used an unsupervised machine learning-based approach to identify subphenotypes. Our method, an extension of the Phe2Vec algorithm, uses tens of thousands of clinical concepts from multiple domains to represent patients' clinical histories to then identify groups of patients with similar presentations. The results indicate that cardiorespiratory presentations are most common (present in 54% of patients) followed by subphenotypes marked (in decreasing order of frequency) by musculoskeletal pain, neuropsychiatric conditions, gastrointestinal symptoms, headache, and fatigue.

Figure 1:. Subphenotype model flowchart

Figure 2:. Subphenotype embeddings of PASC cohort B clinical histories

Figure 3:. Heatmap of incident post-acute diagnoses, cohort B

Summary of clusters by presence of incident PASC-associated diagnoses. To be counted, diagnoses in the respective clusters had to occur in the 28–179 post-acute period following infection and not have been present in the 18 months prior. Cells display proportions of patients in the cluster with the corresponding PASC-associated diagnosis group, and the results of Compact Letter Display (CLD) analysis are represented in superscripts. For a given incident PASC-associated diagnosis group (row), two clusters share the same letter when proportions did not differ significantly (via multiple-testing adjusted chi squared testing) between the two clusters.