Multi-method phenotyping of Long COVID patients using high-dimensional symptom data

Background Long COVID, characterized by symptoms that remain or emerge in the months after infection with COVID-19, has complex and highly variable patient presentations, with myriad seemingly disconnected symptoms. Methods We apply three different machine learning techniques to identify groups of patients with similar symptoms in a large patient-reported symptom dataset with the aim of identifying robust Long COVID phenotypes. Results All three methods produced clinically plausible symptom clusters which are technically valid partitions of the high-dimensional symptom space. However, concordance across methods was low. Some features did recur, such as low-symptom count clusters having the highest average age and lowest proportion of women, and specific recurrent clusters or subclusters across pairs of methods. Conclusions The high sensitivity of observed patient clusters to algorithm choice has implications for other studies reporting Long COVID phenotype clustering, as it suggests that a single method may provide an incomplete or unstable partition of the cohort, particularly in studies with fewer symptoms observed. With the 162 reported symptoms considered here, patient presentations vary smoothly and segmentation, while internally consistent, was not reproducible across methods; this suggests that the complexity of LC symptom presentation may easily be missed by clustering approaches that use insufficient data or overly-simplistic clustering methods. Future work would likely benefit from semi-supervised approaches matching patients to pre-defined phenotypes or diagnoses, or from the inclusion of additional patient data. Overall, our multi-method analysis highlights the importance of assessing clustering robustness and considering the full scope of patient symptoms when evaluating treatments. *Tessa D. Green, Chris McWilliams, and Leonardo de Figueiredo share first authorship.