The study, led by Edith Visser of the Department of Epidemiology at Frisius MC in Leeuwarden, together with colleagues from University Medical Centre Groningen, Radboud University Medical Centre and other Dutch institutions, enrolled adults aged 18 and over who had been referred to post-COVID outpatient clinics at three hospitals in the Netherlands. All participants experienced persistent symptoms following a SARS-CoV-2 infection. In total, 1,155 patients formed the analytic cohort, a population that was 67 percent female with a median age of 50 years. At their initial clinic visit, the median time since infection stood at 24.1 weeks, placing most patients well into the chronic phase of their illness when they were first characterised.

The methodological heart of the study lies in its use of K-means clustering, a workhorse algorithm of unsupervised machine learning. Rather than asking clinicians to sort patients into categories by intuition, the researchers fed the algorithm the pattern of reported post-COVID symptoms recorded at each patient’s first appointment. K-means works by iteratively assigning each data point to the nearest of a predefined number of cluster centres, then recalculating those centres as the mean of the points assigned to them, until the grouping stabilises. The result is a partition of the patient population into groups whose members resemble one another more closely than they resemble anyone else, without any prior assumption about what those groups should look like. The researchers complemented this with multiple correspondence analysis, a related dimensionality-reduction technique suited to categorical data, to visualise and validate the symptom structure underlying the clusters.

Four clusters emerged from the analysis, and their relative sizes were strikingly uneven. The largest group by far, accounting for 47.7 percent of the cohort, was labelled the Respiratory-dominant and Dyspnoea cluster, characterised chiefly by breathing difficulties. The second-largest, at 27.6 percent, was the Fatigue-dominant and Neurocognitive cluster, in which profound tiredness coexisted with cognitive complaints. A third group, the Fatigue-dominant and Multisystem cluster, made up 12.7 percent of patients and carried the broadest symptom burden, spanning multiple organ systems. The smallest, at 11.9 percent, was the Respiratory-dominant and Airway-Systemic cluster. Notably, the authors highlight, only around 12 percent of patients belonged to the group with predominantly systemic symptoms, a finding that challenges assumptions that post-COVID is primarily a multisystem disorder in clinic-referred populations.

Clustering patients is only scientifically useful if the groups mean something clinically, and this is where the study’s second phase becomes important. The researchers examined how patients in each cluster moved through the healthcare system during their initial care. Patients in the two fatigue-dominated clusters were referred significantly more often to occupational therapy, speech and language therapy, mental health services, or a rehabilitation programme than those in the respiratory-dominant clusters, with all comparisons reaching statistical significance at p below 0.01. Patients in the respiratory-dominant groups, by contrast, more frequently received pharmacological treatment, reflecting the different therapeutic logic of managing airway and breathing symptoms compared with the functional and cognitive impairments that dominate fatigue-led illness.

The long-term outlook was assessed after a mean follow-up of 2.5 years, when all surviving participants were invited to complete a digital questionnaire covering their current health status and quality of life. The analysis leaned heavily on patient-reported outcome measures, or PROMs, instruments that capture how patients themselves rate their fatigue, functional capacity, quality of life and psychological state. Among the tools referenced in the study are the Checklist Individual Strength for fatigue, the Hospital Anxiety and Depression Scale, known as HADS, the EuroQol 5-Dimensions questionnaire for health-related quality of life, and the Clinical COPD Questionnaire for respiratory-specific health status. These instruments translate subjective experience into quantifiable scores that can be compared across groups and over time.

The follow-up data delivered the study’s most consequential finding: the fatigue-dominated clusters continued to show a higher disease burden at two and a half years, reflected in worse PROM outcomes compared with the respiratory-dominated clusters. In other words, the symptom pattern identified at the very first clinic visit predicted who would still be struggling years later. Patients whose initial illness was dominated by breathlessness and respiratory complaints fared comparatively better on these self-reported measures, while those in the fatigue and neurocognitive or multisystem groups remained more impaired. This prognostic separation, visible from day one, suggests that a simple symptom-based classification performed at intake could help clinics identify early which patients will need the most intensive and sustained support.

Perhaps the most troubling trend in the data concerns mental health. Anxiety and depression scores, measured with the HADS instrument, increased over time in all four clusters, not merely in the groups with the heaviest physical symptom burden. This universal deterioration cuts against any hopeful narrative of spontaneous recovery across the board and underscores the psychological toll of living with a chronic, poorly understood condition. It also carries practical implications for care planning: mental health support cannot be reserved for the fatigue clusters alone, because rising psychological distress appears to be a shared feature of the post-COVID experience regardless of the dominant physical symptoms.

The study was funded by the Netherlands Organisation for Health Research and Development, ZonMw, under grant 10430302250002, with the funders having no role in study design, data collection, analysis, interpretation, or the decision to publish. Ethical oversight followed Dutch legislation, with the retrospective use of pre-existing clinical data deemed exempt from full review and informed consent waived for that component, while written informed consent was obtained from all respondents to the follow-up questionnaire. The authors declared various competing interests, including trial-related roles for two investigators with pharmaceutical companies outside the submitted work, while the remaining authors reported none.

For clinicians and health system planners, the message of this research is that the post-COVID population, despite its notorious heterogeneity, is not an undifferentiated mass. Distinct symptom-based groups can be identified with a standard clustering algorithm at the point of first assessment, and those groups differ meaningfully in the care they consume and the outcomes they experience. The authors argue that these findings can support more tailored management strategies based on patients’ symptom profiles, enabling effective care planning and earlier access to appropriate interventions. A patient assigned to a fatigue-dominant cluster at intake might be fast-tracked toward rehabilitation and multidisciplinary support, while a respiratory-dominant patient might be directed toward pharmacological management of airway symptoms. As health systems worldwide continue to absorb the long tail of the pandemic, evidence that a data-driven snapshot taken in the first months of illness can forecast the years that follow offers a practical tool for allocating scarce specialist resources where they are needed most.

Subject of Research: Symptom-based clustering of post-COVID patients and its association with healthcare utilisation and long-term health outcomes

Article Title: Symptom-based clustering of post-COVID patients and its association with healthcare utilisation and long-term health outcomes: a multicentre cohort study

Article References: Visser, E., Fokkens, A. S., Koning, K. J., van Geffen, W. H., Visser, A., & Kuijvenhoven, J. C. (2026). Symptom-based clustering of post-COVID patients and its association with healthcare utilisation and long-term health outcomes: a multicentre cohort study. BMC Medicine. https://doi.org/10.1186/s12916-026-05307-8

Image Credits: AI Generated

DOI: 10.1186/s12916-026-05307-8

Keywords: post-COVID, long COVID, K-means clustering, symptom clusters, fatigue, dyspnoea, healthcare utilisation, patient-reported outcome measures, anxiety, depression, cohort study, BMC Medicine