60 - A DATA-DRIVEN MULTIOMIC APPROACH IDENTIFIES SHARED CLINICAL AND BIOLOGICAL SIGNATURES IN MULTISYSTEM INFLAMMATORY SYNDROME IN CHILDREN AND KAWASAKI DISEASE

Background/Aim

Multisystem inflammatory syndrome in children (MIS-C) is a post-infectious hyperinflammatory condition temporally associated with SARS-CoV-2 infection or exposure in children. Initial valuable insights towards the management and treatment of MIS-C have been derived from KD, which it clinically resembles. However, the absence of a standardized diagnostic criteria for MIS-C, coupled with its clinical heterogeneity, have resulted in limited and conflicting data describing the pathobiology driving MIS-C and its connection to KD. Therefore, we aim to systematically characterize and compare children with MIS-C and KD using an unsupervised machine learning approach applied to multi-omic patient data. 

Methods

A standardized set of clinical data and pre-treatment biospecimens from 31 MIS-C patients were collected prospectively at a tertiary center. Corresponding measures were obtained from 35 pre-pandemic KD patients through an associated study. Serum cytokines, soluble cytokine receptors and interferon response gene (IRG) expression were quantified in biospecimens using Luminex and NanoString technologies, respectively. The data collected underwent dimensionality reduction by cross-validated probabilistic principal component analysis (PPCA). Sparsified PPCA scores were subsequently used for patient clustering by Gaussian Mixture Models (GMM). The resulting patient groups were characterized by analyzing patterns in biological profiles, clinical phenotypes and treatment outcomes. 

Results

Sparsified PPCA produced four composite signatures that captured approximately 60% of dataset variation (Figure 1): 1) Elevated IRGs and cytopenias 2) Elevated cytokines and young age, 3) Hyperinflammation and cytokine antagonists, and 4) Platelet and endothelial activation. GMM generated four patient clusters with distinct clinical and biological profiles that corresponded to disease severity, treatment response and patient outcome. The clusters were summarized as unique endotypes: Hyperinflammatory KD, Interferon (IFN)-mediated KD, Mild KD, and KD Shock (Figure 2). Additionally, this patient stratification revealed associations with genetic variants linked to increased cytokine expression. A p-value titration of response variables indicated that the cluster classification outperforms traditional clinical diagnoses, clinical phenotypes and institutional MIS-C case definitions in identifying homogeneous patient groups (Figure 3). 

Conclusion

Data-driven machine learning approaches identified clinically and biologically meaningful patient subgroups in children experiencing post-infectious hyperinflammation. These clusters exhibit clear associations with disease severity and patient outcome, irrespective of diagnosis. This contrasts with data using clinical disease definitions, highlighting the importance of multi-omic patient characterization and suggests that KD and MIS-C fit within the same disease spectrum.