Routine laboratory data are decoupled from symptom phenotype and prognosis in Long COVID: within-cohort variation is...

Long COVID is defined by persistent multi-organ symptoms without an alternative diagnosis. This study evaluated whether routine laboratory data collected under real-world conditions differentiate symptom-defined subgroups within an outpatient cohort and predict the course.

In this single-center cohort study of 564 Long COVID patients (1,452 visits), symptom phenotypes were derived by latent class analysis of 19 organ-system domains at baseline. Structure in 89 routine analytes was sought by k-means clustering, confirmed with alternative algorithms, and repeated after age residualization. Analytes were compared across phenotypes using Kruskal-Wallis tests with false-discovery-rate control and adjustment for confounders. Equivalence testing was performed against reference change values. The course was modeled by count models. Prediction of improvement used a nested cross-validated elastic net with a baseline-independent endpoint.

Clustering showed no convincing structure (silhouette ≤0.09, two to eight clusters); the weak split tracked age (50.5 vs. 38 years) and vanished after residualization. Three age-independent phenotypes (n=503) had no laboratory correlate: none of 89 analytes survived FDR correction, with or without adjustment, or across 50 pseudo-class draws. Equivalence held for 86/89 analytes within ±0.5 SD and 63/70 within the reference change value. Symptom burden declined (median 4 to 2 organ systems, p=2.5 × 10^-19); however, no analyte predicted the course (AUC 0.46).

Routine laboratory variation in this cohort was dominated by age and decoupled from symptom phenotype in both directions. Routine testing in a real-world setting could not provide an age-independent diagnostic, subtyping, or prognostic signature in a cohort of patients suffering from Long COVID.