176 - DATA-DRIVEN CONCORDANCE OF CLINICAL PROFILE BETWEEN SUBSETS OF MIS-C AND PRE-PANDEMIC KD PATIENTS AS DETERMINED BY AN UNSUPERVISED MACHINE LEARNING ALGORITHM

Background: Recent studies have shown that some patients with KD diagnosed before the COVID pandemic have immunological signatures that could be consistent with that of MIS-C. We hypothesized that a clinical profiling tool previously developed by our team to differentiate KD from MIS-C could help us investigate whether any pre-COVID KD patients had a clinical profile consistent with MIS-C.

Methods: We previously developed an unsupervised machine learning algorithm to classify patients on the KD to MIS-C clinical spectrum (along with MIS-C negative COVID controls) into homogenous subgroups based on the degree of similarity of clinical profile at admission (excluding clinical labels and outcomes). In short, we used 93 clinical features and performed a combination of Uniform Manifold Approximation and Projection (UMAP) algorithm for dimensionality reduction and Gaussian Mixture Model (GMM) for spectral projection and clustering. Our algorithm identified 7 distinct clusters in the KD to MIS-C clinical spectrum. As part of this study, we projected the clinical data from 1,956 patients with KD diagnosed at the Hospital for Sick Children in Toronto, Canada, between 2000 and 2019 into the spectral space of our clustering algorithm.

Results: Out of the 7 original clusters, 3 were not represented (less than 10 patients in cluster) in the historical KD cohort: A: COVID with mild inflammation and non-specific systemic findings, G: MIS-C negative COVID controls and E: MIS-C in high-risk patients. The remaining 4 clusters were represented in the historical cohort: F: Severe/high-risk KD (N=700, 36%), B: Low-risk KD (N=574, 30%), D: MIS-C with KD-like phenotype (N=632, 32%) and C: Severe MIS-C (N=39, 2%). Patients in clusters D and C were older (5.1±3.5 vs. 3.7±2.7 years, p=0.001) and more likely to be of southeast Asian ethnicity (8.6% vs. 4.6%, p=0.01). Cluster D were more likely to report respiratory symptoms (39% vs. 30%, p=0.02), while patients in cluster C were more likely to report abdominal symptoms (92% vs. 33%, p=0.001). Patients in cluster D (4.0%) and C (12.1%) were more likely to be admitted to the ICU (1.7%, p=0.001), and patients in cluster C were more likely to be diagnosed with shock (14.7% vs. 1.7%, p=0.001).

Conclusion: A substantial number of patients diagnosed with KD prior to the COVID-19 pandemic have a clinical profile consistent with subgroups of MIS-C patients. These patients have different demographic and clinical profiles than other pre-pandemic KD patients and are at substantially higher risk of shock and LV dysfunction.