Intravenous immunoglobulin resistance in Kawasaki disease patients: prediction using clinical data

Abstract

Background: About 10-20% of Kawasaki disease (KD) patients are resistant to the initial infusion of intravenous immunoglobin (IVIG). The aim of this study was to assess whether IVIG resistance in KD patients could be predicted using standard clinical and laboratory features.

Methods: Data were from two cohorts: a Korean cohort of 7101 KD patients from 2015 to 2017 and a cohort of 649 KD patients from San Diego enrolled from 1998 to 2021. Features included laboratory values, the worst Z-score from the initial echocardiogram or during hospitalization, and the five clinical KD signs at presentation.

Results: Five machine learning models achieved a maximum median AUC of 0.711 [IQR: 0.706-0.72] in the Korean cohort and 0.696 [IQR: 0.609-0.722] in the San Diego cohort during stratified 10-fold cross-validation using significant laboratory features identified from univariate analysis. Adding the Z-score, KD clinical signs, or both did not considerably improve the median AUC in either cohort.

Conclusions: Using commonly measured clinical laboratory data alone or in conjunction with echocardiographic findings and clinical features is not sufficient to predict IVIG resistance. Further attempts to predict IVIG resistance will need to incorporate additional data such as transcriptomics, proteomics, and genetics to achieve meaningful predictive utility.

Impact: We demonstrated that laboratory, echocardiographic, and clinical findings cannot predict intravenous immunoglobin (IVIG) resistance to a clinically meaningful extent using machine learning in a homogenous Asian or ethnically diverse population of patients with Kawasaki disease (KD). Visualizing these features using uniform manifold approximation and projection (UMAP) is an important step to evaluate predictive utility in a qualitative manner. Further attempts to predict IVIG resistance in KD patients will need to incorporate novel biomarkers or other specialized features such as genetic differences or transcriptomics to be clinically useful.

Fig. 1. UMAP embeddings of KD patients for the Korean and San Diego cohorts labeled by IVIG resistance.

a Normalized significant laboratory features only, b normalized significant laboratory features and echocardiographic findings (worst Z-score), c normalized significant laboratory features and significant classical KD signs, and d normalized significant laboratory features, echocardiographic findings (worst Z-score), and significant classical KD signs.

Fig. 2. Performance of models trained on feature sets defined through recursive feature elimination.

a Median AUC from the Korean cohort, b median AUC from the San Diego cohort, c sensitivity from the Korean cohort, and d sensitivity from the San Diego cohort. LR logistic regression, RF random forest, GBM gradient boosting machine, NN neural network.