Long COVID Symptom Clusters: What 793 Patients Taught Us
Machine learning analysis of 793 Long COVID patients revealed 6 distinct symptom clusters with dramatically different prognoses. Mixed mental-physical symptoms had 85% PASC rates vs 51% for pure physical.
Key Statistics
| Participants studied | 793 |
| ML classification accuracy | 77.2% |
| Still symptomatic at 1 year | 85.7% |
The 6 Long COVID Phenotypes
| Cluster | Description | PASC Rate | Implication |
|---|---|---|---|
| Physical + Mental (Cluster 6) | Both physical fatigue/pain AND depression/anxiety severely affected | 85% | Treat both domains simultaneously |
| Triple-Mix (Cluster 2) | Physical, mental, AND sleep all moderately affected | 80% | Start with sleep optimization |
| Mental-Dominant (Cluster 3) | Anxiety, depression, and cognitive symptoms dominate | 66% | PHQ-9/GAD-7 screening essential |
| Sleep-Dominant (Cluster 1) | Poor sleep quality is the core issue | 58% | Sleep study first |
| Physical-Dominant (Cluster 5) | Fatigue, pain, breathlessness dominate | 51% | Better prognosis - focus on pacing |
| Asymptomatic (Cluster 4) | Minimal symptoms across all domains | 1% | Recovery trajectory |
Top 5 Predictors (ML Feature Importance)
- Functional status (PCFS) - 0.0766
- Fatigue score - 0.0720
- Sleep quality (PSQI) - 0.0666
- Initial COVID severity - 0.0655
- Cognitive complaints (PROMIS) - 0.0626
Psychosocial predictors also ranked highly: PHQ-9 Depression (7th), GAD-7 Anxiety (9th), PSS Stress (11th), MSPSS Social Support (13th).
Cognitive Testing Findings
| Cognitive Domain | Finding | P-value |
|---|---|---|
| Visual search (active) | Slower in PASC | p = 0.011-0.013 |
| Visual search (passive) | Slower in PASC | p = 0.003-0.005 |
| Processing speed (digit-symbol) | Slower in PASC | p = 0.037 |
| Sustained attention (CPT) | Slower reaction time | p = 0.040 |
What To Do With This Information
- Identify your cluster: Rate what percentage is physical vs mental vs sleep
- Measure psychosocial factors: Complete PSS-10 (stress) and MSPSS (social support)
- Match treatment to cluster: Physical-dominant: pacing. Mental-dominant: therapy. Sleep-dominant: sleep study.
- Take social support seriously: MSPSS ranked 13th predictor - ahead of BMI
Sources
- Staggs H (2025). Machine learning classification of Long COVID in 793 participants. Psychological Research. DOI: 10.1016/j.psychres.2025.116813