Interpretable web-based machine learning model for predicting intravenous immunoglobulin resistance in Kawasaki disease

Figures

• 

  Fig. 1. Study flowchart. The figure illustrates the study design and workflow, including model development and evaluation

• 

  Fig. 2. Heatmap of Spearman’s variable correlations. Variables with more than 50% missing values were excluded. Multiple imputation was applied to manage missing values for the retained variables. Spearman’s correlation analysis was performed for all variables, and those with a low correlation with IVIG resistance were excluded (if their correlation coefficient exceeded 0.6). The final heatmap displays the correlations among the remaining 40 variables

• 

  Fig. 3. Analysis of model performance and feature importance. a Receiver operating characteristic (ROC) curves for all models predicting intravenous immunoglobulin resistance (IVIGR), with the corresponding area under the curve (AUC) values. b–d Model performance metrics during stepwise feature selection. AUC (b), sensitivity (c), and specificity (d) of the three optimal models (i.e., AdaBoost, Random Forest, and RF Ranger) are plotted against the number of selected features. e–g SHAP summary plots for the three optimal models, illustrating the relationship between actual feature values and SHAP values, as well as their impact on model predictions

• 

  Fig. 4. Scatter plots of feature values and SHAP values in the RF model. a–j Comparison of actual feature values and corresponding SHAP values for the top 10 most important features in the RF model for predicting IVIG resistance (IVIGR). The x-axis represents actual feature values, while the y-axis represents SHAP values. A SHAP value greater than 0 indicates a positive contribution to IVIGR prediction

• 

  Fig. 5. Local interpretation of the RF model using SHAP values. a–d SHAP values were used to interpret model predictions for four patients, by assessing the contribution of each feature to individual predictions. Red bars indicate a positive contribution to intravenous immunoglobulin resistance (IVIGR), while blue bars indicate a negative contribution

• 

  Fig. 6. Web-based IVIG resistance risk prediction and probability distribution in patients with KD. a The IVIG resistance (IVIGR) probability is computed based on 10 key clinical parameters using a web-based prediction model. b The violin plot visualizes the predicted probability distribution of IVIGR across patient groups. Blue indicates IVIG-responsive patients, whereas red represents IVIG-resistant patients