(125) Reproducibility of computable phenotypes for post-acute sequelae of COVID-19 (PASC) in US administrative claims data

Background: Post-acute sequelae of COVID-19 (PASC), also known as long COVID, refers to a constellation of persistent symptoms and health complications following a SARS-CoV-2 infection. Presentation is heterogeneous, and findings from a cohort study (INSIGHT) suggest four distinct phenotypes: (1) cardiac/renal, (2) respiratory/sleep/anxiety, (3) musculoskeletal/nervous, and (4) digestive.

Objectives: We evaluated whether the INSIGHT clusters were transportable to different data sources by applying the topic modeling approach to a large US administrative claims database.

Methods: The cohort included individuals of all ages with at least one diagnosis code for long-COVID (ICD-10-CM: B94.8 with U07.1 in the 365 days prior, or U09.9) between 01 JAN 2021 and 31 AUG 2023, who had 365 days of continuous health plan eligibility in HealthVerity, a US administrative healthcare database including medical and pharmacy claims. Incident health conditions after long-COVID were defined based on first diagnosis with a 365-day washout. Dimension reduction was applied to 137 Clinical Classifications Software Refined (CCSR) categories using Poisson factor analysis. A hierarchical agglomerative clustering method with Euclidean distance calculation and Ward linkage criterion was used to create subphenotypes. Incidence was explored using radar plots.

Results: There were 205,007 individuals meeting all study entry criteria. In general, incident probabilities for CCSR domains were lower than those observed in the INSIGHT cohort; however, the patterns of incident diagnosis were similar in several domains. An exception was chest pain/abnormalities of heartbeat, which was more evenly distributed across topics in the HealthVerity database than what had been previously described. Summarizing results from 4–10 clusters, a solution of 6 clusters was best supported. Two resulting subphenotypes aligned with the INSIGHT cohort: digestive and musculoskeletal/nervous. Other emerging subphenotypes were characterized by (1) allergic reactions/upper respiratory infections, (2) dizziness/headache/sleep-wake disorders, (3) cardiac conditions, and (4) elevations across all subphenotype components explored.

Conclusions: Phenotype clusters for PASC from the INSIGHT cohort were not completely reproduced in HealthVerity despite common emerging themes. This may be attributable to differences in underlying data collection by source, limited capture of initial infection in our data, or to variability among PASC patients. In general, differentiation between subphenotypes was less pronounced in the claims data, suggesting that incorporation of clinical data may be needed to enhance the phenotyping of PASC.