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

Sepsis-induced myocardial dysfunction (SIMD) is a common and heterogeneous complication in patients with sepsis and is associated with increased mortality. This study aimed to identify distinct clinical phenotypes of SIMD using unsupervised deep learning and to develop an optimal short-term survival prediction model based on phenotypic stratification. Data from SIMD patients in the MIMIC-III and MIMIC-IV databases were retrospectively analyzed. An autoencoder was used for feature compression, followed by Uniform Manifold Approximation and Projection (UMAP) and K-means clustering to identify phenotypes, with a novel composite scoring system applied to ensure robust cluster selection. Prognostic differences among phenotypes were evaluated using Kaplan-Meier and Cox regression analyses. XGBoost with SHAP (Shapley Additive Explanations) was used for phenotype prediction and model interpretability. Multi-strategy models (M1-M4) were further constructed to assess the predictive value of phenotypic stratification and determine the optimal modeling strategy for survival prediction. Three clusters (Cluster 0-2) with distinct prognostic profiles were identified. Cluster 2 (high-risk phenotype), characterized by metabolic acidosis and multiorgan dysfunction, showed the highest 90-day mortality (55.1% in the development cohort). SHAP analysis highlighted lactate, bilirubin, coagulation indices, and Glasgow Coma Scale as key drivers of phenotype differentiation. Phenotype-specific modeling (M3) significantly outperformed the global model (M1), achieving a validation AUC of 0.880 and a PR-AUC of 0.863 for Cluster 2. Further feature enhancement in M4 did not yield significant additional benefit, supporting M3 as the optimal modeling strategy. These findings delineate clinically actionable SIMD phenotypes using interpretable unsupervised learning and demonstrate that phenotype-specific modeling markedly improves mortality prediction, particularly for Cluster 2 (high-risk phenotype), underscoring the potential of phenotypic stratification to advance precision critical care in SIMD.