ICU Admission Prediction for Patients With Kawasaki Disease or MIS-C Using Machine Learning

Abstract

Background: Multisystem inflammatory syndrome in children (MIS-C) and Kawasaki disease (KD) show a broad spectrum of clinical severity, from a relatively benign clinical course to requiring admission to the intensive care unit (ICU). With either, clinical deterioration may be rapid and unexpected.

Objectives: The aim of the study was to develop a machine learning (ML) model to predict future ICU admission for patients with KD or MIS-C to augment clinical decision-making.

Methods: We developed a prediction model for ICU admission using 2,539 patients <18 years of age with MIS-C or KD enrolled in the International Kawasaki Disease Registry. Using discrete time-point clinical features and engineered time-series clinical features, we developed predictive snapshot and window ML models with logistic regression, XGBoost, and random forest. Performance was compared between the various iterations of the models.

Results: ML models effectively predicted admission to the ICU within the next 48 hours of the time of prediction. The time-series window-XGBoost model outperformed other models with an AUROC of 0.92 and an area under the precision-recall curve of 0.86. The incorporation of engineered time-series features improved the precision and recall independent of the length of the sampling time window. Higher ferritin level, treatment with anticoagulant or unfractionated heparin, higher C-reactive protein level, and lower platelet count were identified as the most predictive features for positive ICU prediction.

Conclusions: ML algorithms can effectively predict ICU admission for pediatric patients with MIS-C or KD. These models may prompt physicians to pre-emptively implement supportive measures, possibly mitigating the risk of clinical deterioration.

Keywords: Kawasaki disease; clinical deterioration; machine learning; multisystem inflammatory syndrome in children (MIS-C); risk prediction.

Graphical abstract

Figure 1

Schematic Prediction Methodology For 1 Patient With Multiple Prediction Times At prediction time 1 to 3, the ICU admission does not occur within 48 hours of prediction, yielding no ICU admission (outcome = 0). At prediction time 4, the ICU admission does occur within 48 hours of prediction, yielding ICU admission (outcome = 1). Light blue represents the observation window (timeframe of the data considered for prediction). Dark blue represents the prediction time (48 hours after the time of prediction). ICU = intensive care unit.

Central Illustration

Machine Learning Model for ICU Admission in Children With Kawasaki Disease or MIS-C

Figure 2

Machine Learning Models Demonstrate Effective Prediction of Intensive Care Unit Admission XGBoost has outperformed the random forest and logistic regression in both snapshot and window-type models. (A) Receiver-operator characteristic (ROC) curves with corresponding area under ROC (AUROC) of snapshot and window models (XGBoost with the highest AUROC = 0.89 and 0.92, respectively). (B) Precision-recall curves (PRCs) with corresponding area under PRC (PRAUC) of snapshot and window models (XGBoost with the highest PRAUC = 0.81 and 0.86, respectively). (C) Calibration curves of snapshot and window models demonstrate a good agreement between predicted probabilities and observed outcomes. (D) Bar plots of PRAUC (black) and AUROC (pink) metrics for both snapshot and window models. XGBoost has the highest AUROC and PRAUC values relative to random forest and logistic regression in both snapshot and window-type models.

Figure 3

Incorporation of Time-Series Clinical Features Improves Predictive Performances Independent of Sampling Time (A) Receiver-operator characteristic (ROC) curves with corresponding area under ROC (AUROC) of snapshot-XGBoost and window-XGBoost models varying sampling window (1 day, 3 days, and 5 days). (B) Precision-recall curves (PRCs) with corresponding area under PRC (PRAUC) of snapshot-XGBoost and window-XGBoost models varying sampling window (1 day, 3 days, and 5 days). (C) Bar plots of AUROC and PRAUC of snapshot-XGBoost and window-XGBoost models varying sampling window (1 day, 3 days, and 5 days).

Figure 4

Machine Learning Models Reveal the Feature Importance Levels Associated With Admission to the Intensive Care Unit The plots show the 15 most important features measured by SHapley Additive exPlanations (SHAP) with the 3-day XGBoost model. A higher absolute value of SHAP means higher predictive characteristics. (Left) Navy represents discrete time-point clinical features. Yellow represents engineered time-series clinical features. (Right) Blue represents relatively lower feature values. Pink represents relatively higher feature values. IVIG = intravenous immunoglobulin.

Figure 5

Means of the Absolute Value of SHapley Additive exPlanations Follow an Exponential Decrease Trend The plot shows the 50 most important features (out of a total of 1,763 features) measured by SHAP with the 3-day XGBoost model. The 15 red data points signify the “linear” decrease region of the graph. SHAP = SHapley Additive exPlanations.

Figure 6

Intensive Care Unit Admission Prediction for Patients With Kawasaki Disease or Multisystem Inflammatory Syndrome in Children Using Machine Learning A parsimonious model was inducted with 15 features that had a high mean absolute value of SHapley Additive exPlanations. (A) Receiver-operator characteristic (ROC) curves with the corresponding area under ROC (AUROC) of window-XGBoost model using all and the top 15 features (AUROC = 0.92 and 0.89, respectively). (B) Precision-recall curves (PRC) with the corresponding area under PRC (PRAUC) of window-XGBoost model using all and the top 15 features (PRAUC = 0.86 and 0.81, respectively). This analysis demonstrates that these 15 features have a critical impact on predicting the outcome risk.