Phenotype discovery and mortality prediction in sepsis-induced myocardial dysfunction: a deep learning and stratified modeling approach

Fig. 1. Three-dimensional visualization of SIMD clusters in the development and independent temporal cohorts. (A) Development cohort; (B) Independent temporal cohort. Each point represents an individual patient projected onto the three-dimensional UMAP embedding space derived from the learned latent representation. Colors indicate cluster assignments obtained using K-means clustering. Black stars denote cluster centroids. The clustering framework and parameter settings were applied consistently across both cohorts

Fig. 2. Mortality and survival analysis across clustering subgroups in the development and independent temporal cohorts. (A, B) Bar plots showing 28-day and 90-day all-cause mortality rates across clusters in the development cohort (A) and independent temporal cohort (B). (A1, B1) Kaplan–Meier survival curves for 28-day mortality; (A2, B2) Kaplan–Meier curves for 90-day mortality. Shaded areas represent 95% confidence intervals. Numbers at risk are displayed below each plot. Across both cohorts, Cluster 2 showed consistently lower survival probabilities compared with Clusters 0 and 1 (log-rank test, P < 0.001)

Fig. 3. SHAP-based interpretation of feature importance for cluster assignment in the development and independent temporal cohorts. (A) SHAP results in the development cohort; (B) SHAP results in the independent temporal cohort. Heatmaps display the mean absolute SHAP values of each feature across clusters, indicating their relative contribution to model predictions of cluster assignment. Panels A1 and B1 present global feature importance rankings across all clusters. Panels A2–A4 and B2–B4 show cluster-specific feature importance, highlighting variables most strongly associated with each cluster in the trained classification model

Fig. 4. Local SHAP explanations for representative correctly predicted samples from each cluster. Panels show SHAP force plots for representative patients in the development cohort (A) and the independent temporal cohort (B). A1–A3 and B1–B3 correspond to Cluster 0, Cluster 1, and Cluster 2, respectively. Red bars indicate features contributing positively to the model output for the predicted class, whereas blue bars indicate features contributing negatively. The plots illustrate how individual feature values influence model predictions at the patient level

Fig. 5. Comprehensive evaluation of the cluster-specific predictive model for Cluster 2 (high-risk phenotype). Panels present the receiver operating characteristic (ROC) curve, precision–recall (PR) curve, decision curve analysis, confusion matrix, calibration curve, risk stratification, Kolmogorov–Smirnov (KS) curve, predicted probability distribution, and clinical impact curve for Cluster 2. Model performance is shown for internal cross-validation (development cohort) and the independent temporal cohort. Shaded areas represent ± 1 standard deviation where applicable. These analyses collectively evaluate model discrimination, calibration, clinical utility, and risk stratification ability

Fig. 6. SHAP-based interpretation of the predictive model for Cluster 2 (high-risk phenotype). ()) Global feature importance based on mean absolute SHAP values. (B) SHAP beeswarm plot showing the distribution of feature contributions across all patients. (C) SHAP decision plot illustrating cumulative feature effects on model predictions for representative samples. (D) SHAP waterfall plot for a representative survivor. (E) SHAP waterfall plot for a representative non-survivor. SHAP analysis identifies features most strongly associated with model predictions and illustrates their directional contributions to predicted risk. Supplementary figures provide analogous analyses for other clusters and the independent temporal cohort