Multimodal deep learning for risk stratification of imaging subtypes in pneumonia-associated sepsis
Pneumonia-associated sepsis shows substantial heterogeneity in lung injury and clinical outcomes. We analyzed a retrospective multicenter cohort of 1,088 patients and a next-generation sequencing (NGS) subgroup of 94 patients to evaluate pulmonary radiomic phenotypes, host body composition, and reported lower respiratory tract pathogen profiles. After ComBat harmonization, principal component analysis and K -means clustering were used to derive four imaging subtypes in the internal training cohort, which were transferred to validation cohorts and NGS subgroup. The subtypes showed distinct clinical outcome patterns and selected associations with body-composition and inflammatory measures in adjusted analyses. NGS report-derived pathogen-profile findings remained exploratory, with no pathogen-level differences surviving false discovery rate (FDR) correction. A multimodal transformer integrating clinical-laboratory, radiomic, and body-composition features showed numerically higher external-validation area under the receiver operating characteristic curves (AUCs) than selected conventional models. These findings provide a cross-scale framework for clinically interpretable risk stratification in pneumonia-associated sepsis.
Authors
- 李郎俊
- Guangjun Fan (ORCID: https://orcid.org/0000-0001-8532-1660)
- Marcin Grzegorzek (ORCID: https://orcid.org/0000-0003-4877-8287)
- Tianming Du (ORCID: https://orcid.org/0000-0001-7671-6513)
- Hongbo Wang (ORCID: https://orcid.org/0009-0006-1625-3629)
- yin xiong
- Jianqiong Chen
- Chen Li
- Yuanheng Zhang
Institutions
- Second Affiliated Hospital of Dalian Medical University (CN)
- The Affiliated Yongchuan Hospital of Chongqing Medical University (CN)
- Chongqing Medical University (CN)
- China Medical University (CN)
- Northeastern University (CN)
- University of Lübeck (DE)
Publication Details
- Journal
- iScience
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1016/j.isci.2026.118304
- Primary Topic
- Radiomics and Machine Learning in Medical Imaging
- Type
- article
- Field-Weighted Citation Impact
- 0.00