A machine learning approach to predict intravenous immunoglobulin resistance in Kawasaki disease patients: A study based on a Southeast China population

Abstract

Kawasaki disease is the leading cause of pediatric acquired heart disease. Coronary artery abnormalities are the main complication of Kawasaki disease. Kawasaki disease patients with intravenous immunoglobulin resistance are at a greater risk of developing coronary artery abnormalities. Several scoring models have been established to predict resistance to intravenous immunoglobulin, but clinicians usually do not apply those models in patients because of their poor performance. To find a better model, we retrospectively collected data including 753 observations and 82 variables. A total of 644 observations were included in the analysis, and 124 of the patients observed were intravenous immunoglobulin resistant (19.25%). We considered 7 different linear and nonlinear machine learning algorithms, including logistic regression (L1 and L1 regularized), decision tree, random forest, Ada-Boost, gradient boosting machine (GBM), and lightGBM, to predict the class of intravenous immunoglobulin resistance (binary classification). Data from patients who were discharged before Sep 2018 were included in the training set (n = 497), while all the data collected after 9/1/2018 were included in the test set (n = 147). We used the area under the ROC curve, accuracy, sensitivity, and specificity to evaluate the performances of each model. The gradient GBM had the best performance (area under the ROC curve 0.7423, accuracy 0.8844, sensitivity 0.3043, specificity 0.9919). Additionally, the feature importance was evaluated with SHapley Additive exPlanation (SHAP) values, and the clinical utility was assessed with decision curve analysis. We also compared our model with the Kobayashi score, Egami score, Formosa score and Kawamura score. Our machine learning model outperformed all of the aforementioned four scoring models. Our study demonstrates a novel and robust machine learning method to predict intravenous immunoglobulin resistance in Kawasaki disease patients. We believe this approach could be implemented in an electronic health record system as a form of clinical decision support in the near future.

Predicting bloodstream infection outcome using machine learning

“…We implemented the gradient-boosting predictor trained with the LightGBM 40 Python package. LightGBM has shown effectiveness on clinical and patient tabular data in particular, and was adopted by many recently published models 41 – 46 . Missing values were inherently handled by the LightGBM predictor 40 , 47 , 48 .…”

Predicting bloodstream infection outcome using machine learning

“…We implemented the gradient-boosting predictor trained with the LightGBM [37] Python package. LightGBM has shown effectiveness on clinical and patient tabular data in particular, and was adopted by many recently published models [38][39][40][41][42][43]. Missing values were inherently handled by the LightGBM predictor [37,44,45].…”

A Simple Scoring Model Based on Machine Learning Predicts Intravenous Immunoglobulin Resistance in Kawasaki Disease

“…To establish a more reliable and simple scoring system for the prediction of IVIG resistance in KD patients, an alternative approach using large data repositories is required. Recently the developed machine learning approach has shown great potential for assisting the clinical diagnosis and predicting outcomes [14][15][16][17][18][19]. Two recent studies applied machine learning to predict IVIG resistance in KD patients, and con rmed its usefulness [14,15].…”

“…Recently the developed machine learning approach has shown great potential for assisting the clinical diagnosis and predicting outcomes [14][15][16][17][18][19]. Two recent studies applied machine learning to predict IVIG resistance in KD patients, and con rmed its usefulness [14,15]. However, there were several limitations in both studies including a limited number of KD patients (n=98) in a single institute in one study [14], and a relatively large number of KD patients (n=497) with two different IVIG protocols in the other [15].…”

“…Two recent studies applied machine learning to predict IVIG resistance in KD patients, and con rmed its usefulness [14,15]. However, there were several limitations in both studies including a limited number of KD patients (n=98) in a single institute in one study [14], and a relatively large number of KD patients (n=497) with two different IVIG protocols in the other [15]. Moreover, in clinical practice, even if machine learning has a high degree of accuracy, a simple scoring system is more convenient for risk-strati ed treatment.…”