Generalisable long COVID subtypes: findings from the NIH N3C and RECOVER programmes

Area of Science:

• Computational biology
• Medical informatics
• Public health

Background:

• Post-acute sequelae of SARS-CoV-2 infection (PASC), or long COVID, presents a complex challenge for clinical management due to its diverse manifestations and computational analysis difficulties.
• The generalizability of machine learning models for predicting COVID-19 outcomes remains largely untested.

Purpose of the Study:

• To develop a computational method for modeling PASC phenotypes using electronic healthcare records (EHRs).
• To assess patient phenotypic similarity using semantic analysis and unsupervised machine learning for patient stratification.

Main Methods:

• Utilized electronic healthcare records (EHRs) to computationally model PASC phenotype data.
• Developed a nonlinear semantic similarity function to map phenotypic abnormalities to pairwise patient similarity.
• Applied unsupervised machine learning for clustering patients based on semantic similarity.

Main Results:

• Identified six distinct clusters of PASC patients with unique profiles of pulmonary, neuropsychiatric, and cardiovascular abnormalities.
• Discovered a cluster associated with severe manifestations and increased mortality, linked to pre-existing conditions and acute COVID-19 severity.
• Demonstrated generalizability of these clusters across different hospital systems by assigning new patients based on maximum semantic similarity.

Conclusions:

• Semantic phenotypic clustering offers a robust foundation for stratifying PASC patients into subgroups.
• Facilitates the design of natural history studies and clinical trials for PASC.
• Enables precision medicine approaches for managing long COVID.