Abstract
Background: Kawasaki disease (KD) is an acute febrile illness of unknown etiology. The most common and alarming complication is coronary artery aneurysm. High dose immunoglobulin therapy (IVIG) is effective in resolving inflammation and reducing the incidence of aneurysm. However, 10-20% of KD children have persistent or recurrent fever after IVIG, remaining high risks of aneurysm. We evaluated which factor is most strongly associated with patient unresponsiveness to IVIG. Methods:We retrospectively reviewed clinical records of 53 KD patients who had been treated with IVIG. They were divided into two groups; 39 patients who responded to IVIG at a dose of 1 g/kg or 2 g/kg (responder group), and 14 who failed to respond to escalated doses, requiring additional treatments (non-responder group). We analyzed a total of 16 variables, including demographic, clinical, and laboratory data, as well as three scoring systems (Gunma, Kurume and Tenri scores). To determine how accurately each variable predicts patient unresponsiveness to IVIG, we calculated the standardized distance between the groups（Δ）, the areas under the ROC curve (AUC),and univariate logistic regression statistics (Wald P). The Tenri score was obtained using logistic stepwise selection based on the likelihood. We then divided the responder group into two subgroups, according to whether patients responded to 1 g/kg IVIG or required a further 2 g/kg dose. Each subgroup was analyzed using the same methods. Results: The Gunma model yielded values for Δand AUC, and a Wald P value, similar to those of the Tenri model, and achieving better results than the Kurume score or other variables. Since the Gunma model was developed and validated using different data sets, it appears to be the most reliable predictor of unresponsiveness to IVIG. However, whereas the Tenri model has four variables, the Gunma model has seven, including the duration of illness before treatment, which reflects severity of the disease (confounding variable). Ockham's razor suggests that a better model could be created using fewer variables. In regard to the distinction between the two subgroups of responders, the Tenri model attained statistical significance, but the Gunma model gave a non-significant result. Conclusion: The Gunma score was able to predict patient unresponsiveness to IVIG more accurately than the Kurume score or other variables, giving comparable results to those of the Tenri score. Nevertheless, the ideal prediction model, with an optimal combination of variables, awaits development using larger datasets.
