Algorithm dependence of patient phenotypes in Long COVID: a patient-led, multi-method clustering of 6031 patients...

Long COVID, characterized by symptoms that remain or emerge in the months after acute COVID-19 infection, is a multisystemic condition with highly variable patient presentations. Phenotyping studies have reported divergent symptom clusters, increasingly used to design trials and interpret biomarker data. However, robustness of these clusters across analytic methods remains uncertain.

We analyzed data from 6 031 adults with ≥ 90 days of illness from a patient-led international survey. Participants reported presence/absence of 162 symptoms, post-exertional malaise severity and demographics. We applied three unsupervised machine learning approaches to the same symptom matrix, evaluating the resulting clusterings for concordance, robustness to subsampling, and relationship to symptom burden, post-exertional malaise severity, age and gender.

Each method produced clinically plausible symptom clusters, but concordance across methods was low. All three approaches identified a high-symptom-burden group enriched for post-exertional malaise severity, and lower-symptom-burden groups with older mean age and a lower proportion of women. Symptom count consistently correlated with higher post-exertional malaise severity and a greater proportion of women. Manifold analysis revealed that the overall symptom space was largely continuous, lacking clear cluster boundaries.

The strong dependence of patient clusters on algorithm choice suggests that single-method Long COVID phenotyping may produce incomplete or unstable subgroup definitions. Clustering methods may impose artificial boundaries on a smoothly varying symptom landscape, especially in studies capturing fewer symptoms. Phenotyping efforts should assess clustering robustness and avoid overinterpreting single-method results. Our multi-method analysis highlights the importance of considering the full breadth of patient symptoms when evaluating treatments.