Identifying Clinical Subtypes of Long COVID: An Unsupervised Machine Learning Approach

ABSTRACT

Objectives: The symptoms of patients with post-acute COVID-19 syndrome are heterogenous, impact multiple systems, and are often non-specific. To better understand the symptomatic profile of this population, this study used real-world data and unsupervised machine learning techniques to identify distinct groupings of long COVID patients. Method(s) Children/adolescents (age 0-17) and adults (age 18-64 and >=65) with >=2 primary diagnoses for U09.9 "Post COVID-19 condition" from 10/01/2021 (ICD-10 code introduction) until 03/31/2022 were selected from Optum's de-identified Clinformatics Data Mart Database, with the first diagnosis deemed index. Included patients had >=1 diagnosis for COVID-19 at least 4 weeks before index and continuous enrollment during the 12 months prior to index. Diagnoses recorded +/-2 weeks from index that were not present prior to the initial COVID-19 diagnosis were captured and used as patient features for k-means clustering. Final cluster assignments were selected based on silhouette coefficient and clinical relevancy of groupings. Result(s) 3,587 patients met eligibility criteria, yielding three clusters. Concurrent symptom domains surrounding index included breathing, fatigue, pain, cognitive, and cardiovascular diagnoses. The first cluster (N=2,578, 71.8%) was characterized by patients with only a single symptom domain (33% breathing, 33% cardiovascular, 20% fatigue, 11% cognitive). The second cluster (N=651, 18.1%) all presented with breathing symptoms accompanied by one additional domain (cardiovascular 40%, fatigue 28%, pain 18%). The final cluster (N=358, 9.9%) experienced breathing symptoms accompanied by two additional domains (fatigue and cardiovascular 34%, cardiovascular and cognitive 34%). Cluster 3 was slightly older than clusters 1 or 2 (mean age 66 vs. 58 years, respectively). Conclusion(s) Unsupervised machine learning identified distinct groups of long COVID patients, which may help inform multidisciplinary care needs. Our analysis suggests that many patients with long COVID may experience symptoms from only a single domain, and multi-system illness may generally include breathing complications accompanied by fatigue and/or cardiovascular complications.Copyright © 2023