Early recognition of sepsis-associated acute kidney injury based on BioBERT-BPNN multimodal clinical data analysis
- 1Shandong Public Health Clinical Center, Shandong University, Shandong, China
Abstract
Background: Sepsis-associated acute kidney injury (SA-AKI) is a common and serious complication in intensive care units, marked by high incidence, rapid disease progression, and poor clinical outcomes. Due to its complex pathophysiology and nonspecific clinical manifestations, early recognition of SA-AKI remains highly challenging. Methods: This study aims to achieve early recognition of SA-AKI using clinical records of sepsis patients. A dual-center dataset of sepsis patients is constructed by integrating 1,000 cases from the publicly available MIMIC-IV database and 400 cases from the electronic medical records system of Shandong Public Health Clinical Center. Multimodal data are generated for each sepsis case by extracting 33 risk factors from multi-source clinical information. A BioBERT-BPNN multimodal model is developed, in which the BioBERT module extracts semantic features from textual data, the BPNN module captures nonlinear features from numerical data, and MLP module is applied for multimodal feature fusion and classification. Results: The proposed model demonstrates stable and reliable performance for early recognition of SA-AKI on the MIMIC dataset and exhibits good cross-center generalizability and robustness on the external Hospital dataset. Dual-center data enhances the model’s adaptability to local cases and improves its clinical applicability. Moreover, multimodal data have complementary value, the multimodal model outperforms single-modality models. Conclusion: The proposed BioBERT-BPNN multimodal model effectively integrates textual and numerical information and improves the performance of early SA-AKI recognition. Leveraging dual-center multi-source heterogeneous clinical data enhances its applicability in real-world clinical settings. This study could provide valuable reference for timely intervention, optimized decision-making, and improved outcomes in sepsis patients at risk of AKI, holding important clinical research value.
Related
- Understanding the processes involved in sepsis recognition across clinical contexts to guide individualised management: A phenomenographic study
- Replacing EEG with peripheral physiological signals for driving behavior recognition: a physiology-informed multimodal approach
- Lysosomal pH dysregulation as a candidate signal-amplifying hub in sepsis-associated lung injury
- CUBE: Multimodal Representation Learning Reveals Biological Structure Across Histomorphology, Spatial Protein Phenotypes, and Transcriptome-Associated Signals
- Alpha-7 nicotinic acetylcholine receptor activation preserves podocyte integrity during sepsis-associated kidney injury
BibTeX
@article{chen2026early,
title = {{Early recognition of sepsis-associated acute kidney injury based on BioBERT-BPNN multimodal clinical data analysis}},
author = {Yufeng Chen and Hongyang Bian and Yuye Li},
journal = {PLOS One},
year = {2026},
month = sep,
doi = {10.1371/journal.pone.0357602},
url = {https://doi.org/10.1371/journal.pone.0357602},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Early recognition of sepsis-associated acute kidney injury based on BioBERT-BPNN multimodal clinical data analysis
AU - Yufeng Chen
AU - Hongyang Bian
AU - Yuye Li
PY - 2026
DA - 2026/09/11
JO - PLOS One
DO - 10.1371/journal.pone.0357602
UR - https://doi.org/10.1371/journal.pone.0357602
AB - Background: Sepsis-associated acute kidney injury (SA-AKI) is a common and serious complication in intensive care units, marked by high incidence, rapid disease progression, and poor clinical outcomes. Due to its complex pathophysiology and nonspecific clinical manifestations, early recognition of SA-AKI remains highly challenging. Methods: This study aims to achieve early recognition of SA-AKI using clinical records of sepsis patients. A dual-center dataset of sepsis patients is constructed by integrating 1,000 cases from the publicly available MIMIC-IV database and 400 cases from the electronic medical records system of Shandong Public Health Clinical Center. Multimodal data are generated for each sepsis case by extracting 33 risk factors from multi-source clinical information. A BioBERT-BPNN multimodal model is developed, in which the BioBERT module extracts semantic features from textual data, the BPNN module captures nonlinear features from numerical data, and MLP module is applied for multimodal feature fusion and classification. Results: The proposed model demonstrates stable and reliable performance for early recognition of SA-AKI on the MIMIC dataset and exhibits good cross-center generalizability and robustness on the external Hospital dataset. Dual-center data enhances the model’s adaptability to local cases and improves its clinical applicability. Moreover, multimodal data have complementary value, the multimodal model outperforms single-modality models. Conclusion: The proposed BioBERT-BPNN multimodal model effectively integrates textual and numerical information and improves the performance of early SA-AKI recognition. Leveraging dual-center multi-source heterogeneous clinical data enhances its applicability in real-world clinical settings. This study could provide valuable reference for timely intervention, optimized decision-making, and improved outcomes in sepsis patients at risk of AKI, holding important clinical research value.
SN - 1932-6203
ER -