Deriving reproducible sepsis clinical subphenotypes through multimodal data integration framework

Sepsis remains a major global health challenge, with profound heterogeneity limiting effective therapies. Data-driven subphenotyping has shown promise for precision treatment, but most existing approaches rely on unimodal or snapshot data with limited reproducibility. While multimodal data may improve patient clustering and treatment-effect estimation, study that integrate multimodal data into sepsis phenotyping remains lacking. To fill this gap, we developed the Multimodal Integration Enabling Novel and Reproducible Subphenotypes (MINERS) framework using 55,940 sepsis patients from China and the United States. The MINERS framework comprises three modules: Discovery, Transfer, and Validation, that enable identification of sepsis subphenotypes from multimodal data, early subphenotype prediction across cohorts, and evaluation of reproducibility and clinical utility. Through MINERS, five data modalities were integrated into a 4945-dimensional patient representation, from which four reproducible subphenotypes were identified through embedding and clustering. The resulting subphenotypes demonstrated distinct clinical trajectories, prognostic patterns, and treatment-response associations, which were further validated across international cohorts. Compared with existing approaches, MINERS-derived subphenotypes more comprehensively capture longitudinal trajectory heterogeneity and provide a transferable and extensible framework for reproducible subphenotyping.