Data-driven clustering identifies features distinguishing multisystem inflammatory syndrome from acute COVID-19 in...

Area of Science:

• Pediatric infectious diseases
• Critical care medicine
• Epidemiology

Background:

• Current consensus criteria for Multisystem Inflammatory Syndrome in Children (MIS-C) prioritize sensitivity, potentially including cases of acute COVID-19 pneumonia.
• Accurate differentiation between MIS-C and severe COVID-19 is crucial for appropriate patient management.

Purpose of the Study:

• To apply an unsupervised clustering approach to identify distinct phenotypes within hospitalized children with COVID-19-related illness.
• To evaluate the accuracy of clinical labels for MIS-C and identify potential misclassifications.

Main Methods:

• Unsupervised clustering was performed on data from 1,526 pediatric patients hospitalized with COVID-19-related illness.
• Clinical features and diagnostic labels were analyzed to identify distinct patient clusters.
• Recursive feature elimination was used to pinpoint characteristics differentiating potentially misclassified cases.

Main Results:

• Three distinct patient clusters emerged based on 46 clinical features.
• Cluster 1 (92% labeled MIS-C) predominantly comprised previously healthy children with cardiovascular/mucocutaneous involvement and negative SARS-CoV-2 PCR.
• Cluster 2 (27% labeled MIS-C) frequently had pre-existing conditions, respiratory involvement, positive PCR, and chest radiograph infiltrates, distinguishing them from Cluster 1.
• Cluster 3 represented younger, PCR-positive patients with less inflammation.

Conclusions:

• Data-driven clustering successfully identified a phenotype highly likely to be MIS-C.
• A separate cluster of patients, characterized by pulmonary infiltrates and positive PCR, was more likely to represent acute severe COVID-19 pneumonia.
• Clinician-labeled MIS-C cases within this latter group may represent misclassifications, suggesting a need to refine MIS-C diagnostic criteria.