Fig. 1: Comprehensive framework of our study.

Data acquisition: three public cohorts (EHRs) from MIMIC-III, MIMIC-IV, and eICU-CRD, a local cohort from the First Affiliated Hospital of Chongqing Medical University, and transcriptomic data from the Ningbo Medical Center Lihuili Hospital, GEO, KEGG, GeneCards, and DisGeNET were used for model development. Prognostic prediction: predictors derived from medical practice and laboratory tests were used for model development. SepsisFormer consisted of a domain-adaptive generator that dealt with multi-center data distribution differences and class imbalance, an integrated transformer that extracted information from the EHRs of sepsis patients into fixed-length semantic vectors, and a multilayer perceptron that make predictions. Explainability analysis: the efficiency of heterogeneous markers, including seven coagulation-inflammatory markers and patient age, was explained through EHR-level, transcriptomic-level, and model-level analysis. Heterogeneity assessment: 1, an automated risk stratification tool (SMART), assessed the risk level of sepsis patients based on eight routine medical measurements. 2, Subphenotypes classified sepsis patients as CIS1 and CIS2 via unsupervised methods. 3, Heterogeneity of anticoagulant treatment (e.g., Heparin) effects was evaluated in subgroups with distinct subphenotypes and risk levels. 4, Sepsis management tools. Created in BioRender. Niu, B. (2025) https://BioRender.com/8py5ibd. EHR Electronic health record, XAI Explainable artificial intelligence, CIS1 Coagulation-inflammatory subphenotype 1, and CIS2 Coagulation–inflammatory subphenotype 2.