Background: Kawasaki disease (KD) is an acute, febrile, vasculitis affecting pediatric patients which may results in the development of coronary artery aneurysms and have lifelong health consequences for children. In the acute phase of the disease, children with KD are treated with intravenous immunoglobulin (IVIG) which is provided over 12-18 hours. A substantial portion of patients fail to experience defervescence with the initial IVIG treatment and require secondary treatment with either IVIG or steroids. Longer durations of fever are strongly associated with increased risk of coronary complications and the risk could be substantially reduced by developing a predictive model to identify patients unlikely to respond to initial IVIG treatment and providing them with more intensive therapy earlier in the course of the disease
Method: Core temperature was collected at irregular time points following the IVIG therapy initiation. Stratified by the outcome, we first modelled the empirical dynamics of the post-IVIG temperature using functional principal component with Principal Analysis by Conditional Estimation (PACE) algorithm to accommodate the sparse data. This method allowed us to represent each patient’s temperature longitudinal data as a smooth curve or functional data. Next, classification using k-nearest neighborhood models for function data were applied to predict the outcome variables using the patient-specific curves. We applied two distant metrics – the maximum and the absolute difference – to quantify patients’ similarities using the smooth curves as well as their corresponding velocity and acceleration during the first 18 hours. Data were 75-25 split to training and test data. Six k-nearest neighborhood models were trained and combined using the Brier-score optimal ensemble method proposed in Fuchs et al. (2015). The model performance was evaluated in terms of misclassification rates using the test data.
Results: The study included 363 patients with a median (IQR) follow-up of 25 [15 – 44] hours. There were in total 182 (50.1%),123 (33.9%) and 58 (16.0%) responders, partial non-responders and non-responders, respectively. All responders responded to the first IVIG infusion. Among the 123 partial non-responders, 74 (60.2%) responded to the second IVIG infusion, 15 (12.2%) responded to steroid treatments and 34 (27.6%) discontinued without further treatment. Among the 58 non-responders, 23 (36.7%) responded to the second IVIG infusion, 25 (43.1%) responded to steroid treatments and 10 (17.2%) discontinued without further treatment.
The k-nearest neighborhood method suggested the temperature and its rate of change are predictive to a patient’s IVIG response with an overall correct classification rate of 77% for the test data. The model also achieves an overall correct classification rate of 82% in predicting if a patient respond to the first, second or steroid treatment.
Discussion: We were able to accurately predict failure to respond to IVIG and response to secondary line of treatment using machine learning. This finding could help us individualize the care of children with KD by more aggressively treated high-risk patients and targeting second line therapy.
PRECISION MEDICINE & DRUG DISCOVERY
Author: Chun Po Fan
Coauthor(s): Chun Po Fan, Mallory L. Downie, Brigitte Mueller, Tanveer H. Collins, Mathew Mathew, Brian W. McCrindle, Cedric Manlhiot
Status: Completed Work