Machine learning-based mortality prediction for pediatric fulminant myocarditis using cytokine profiles

Fig. 1. Flow diagram of patient selection and study cohort classification. All consecutive pediatric fulminant myocarditis cases identified during the study period were included in the final analysis.

Fig. 2. Hierarchical clustering heatmap. The top dendrogram illustrates the clustering of patients, while the side dendrogram displays the clustering of clinical characteristics and cytokines. The right side indicates the two groups (survival and mortality) analyzed. In the heatmap, each column represents an individual patient (color-coded by outcome: green for survival, orange for mortality), and each row represents a distinct clinical characteristic or cytokine, organized by expression similarity. Heatmap colors indicate cytokine expression intensities. BW, body weight; BSA, body surface area; HT, height; LVEF, left ventricular ejection fraction; BNP, B-type natriuretic peptide; CK-MB, creatine kinase-MB.

Fig. 3. Classification performance of the evaluated model assessed by repeated 3-fold cross-validation with 100 repeats. (A) Receiver operating characteristic (ROC) curve obtained by averaging the ROC curves computed from fold-averaged predictions across 100 repeats of 3-fold cross-validation. The dashed diagonal line indicates chance-level discrimination. (B) Precision–recall (PR) curve obtained by averaging the PR curves computed from fold-averaged predictions across 100 repeats of 3-fold cross-validation. The dashed horizontal line denotes the baseline precision corresponding to the prevalence of the positive class (0.33). (C) Calibration plot based on bootstrap resampling, showing the apparent and bias-corrected calibration curves. The diagonal line represents perfect calibration.

Fig. 4. Bootstrap distributions of variable importance in projection (VIP) scores derived from PLS discriminant analysis (PLS-DA). Violin plots show the distribution of VIP scores across 2,000 bootstrap resamples for each variable, with embedded boxplots indicating the median and interquartile range. Variables are highlighted in orange when the lower bound of the bias-corrected and accelerated (BCa) bootstrap confidence interval exceeds 1, indicating stable importance across resamples, whereas variables shown in grey satisfy only the criterion of mean VIP > 1. CK-MB, creatine kinase-MB; Lac, lactate.

Fig. 5. Violin plots showing paired bootstrap distributions of performance differences between models, evaluated using out-of-bag predictions. For each bootstrap resample, performance metrics were computed separately for each model, and pairwise differences were calculated as Model A minus Model B. Figure shows differences between the combined model and the clinical-only model (All − Clinical). Within each panel, differences in accuracy, AUROC, and AUPRC are shown. Boxplots indicate the median and interquartile range, and the dashed horizontal line denotes zero difference.