Generalizable Long COVID Subtypes: Findings from the NIH N3C and RECOVER Programs

Area of Science:

• Computational biology and bioinformatics
• Medical informatics and data science
• Infectious disease epidemiology

Background:

• Post-acute sequelae of SARS-CoV-2 infection (PASC), or long COVID, presents a complex and heterogeneous clinical picture, hindering precise patient management.
• The natural history and computational analysis of PASC manifestations are challenging due to the wide range of symptoms and limited understanding.
• Generalizability of machine learning models for COVID-19 outcomes has been insufficiently tested, necessitating robust and validated approaches.

Approach:

• Developed a computational method to model PASC phenotype data derived from electronic healthcare records (EHRs).
• Employed semantic similarity to assess pairwise phenotypic similarity between patients, creating a nonlinear similarity function.
• Utilized k-means clustering on the patient similarity matrix to identify distinct PASC patient subgroups based on phenotypic abnormalities.

Key Points:

• Identified six distinct clusters of PASC patients, each exhibiting unique profiles of phenotypic abnormalities.
• Demonstrated significant associations between cluster membership, pre-existing conditions, and acute COVID-19 severity.
• Found two clusters associated with severe PASC manifestations and increased mortality, consistent across different healthcare systems.

Conclusions:

• Semantic phenotypic clustering offers a robust method for stratifying PASC patients into clinically relevant subgroups.
• The identified clusters are generalizable across hospital systems, supporting their utility in diverse clinical settings.
• This approach provides a foundation for natural history studies and targeted therapeutic interventions for PASC.