Clustering the Chaos: A Computational Analysis Investigating the Heterogeneity of Kawasaki Disease

Abstract

Kawasaki disease (KD) is a rare autoimmune disorder affecting nearly 20 in every 100,000 U.S. children aged 5 years and under. KD causes a harmful immune response towards blood vessels, which can lead to coronary artery aneurysm (CAA). The causes and mechanism of action of KD remain largely a mystery, despite decades of etiological research posing a range of biological, environmental, and genetic causative factors. Due to this lack of information, there are no specific diagnostic tests for the disease, and it is diagnosed only by the presentation of a combination of symptoms. This represents a major deficiency, considering that timely and specific diagnosis is critical for preventing life-threatening complications. However, original research has suggested the existence of several subtypes with distinct clinical features under the broad categorization of KD. These subtypes associate specific variables of KD patients to 4 disease subtypes through hierarchical clustering on principal components (HCPC). Yet, my findings in the process of reproducing this work demonstrated significant shortcomings in the previous methods. Upon extensive testing, the HCPC method proved too sensitive to produce convincing clusters using the available variables. Therefore, future KD research must not rely on HCPC, given its demonstrated instability and risk of misleading subtyping. I thus employ unsupervised k-means clustering and produce 3 putative distinctive, characteristic patient subgroup clusters, which appear markedly less sensitive, providing a more stable framework for following KD subgrouping research. These results suggest that KD may be a syndrome with several distinct variants, requiring targeted care approaches.