Related Experiment Video
Updated: Aug 5, 2026
07:50
A Metadata Extraction Approach for Clinical Case Reports to Enable Advanced Understanding of Biomedical Concepts
Published on: September 20, 2018
A parallel dual-stream state-space module for reliable and efficient biomedical relation extraction
Yaxun Jia1, Zhu Yuan2, Lian Zhu1
1Department of Radiation Oncology, Shanghai East Hospital, Tongji University School of Medicine, Shanghai, China.
Plos Computational Biology
|July 30, 2026
Summary
Automated drug-drug interaction (DDI) extraction is improved by DuSSM, a new framework that separates local patterns from semantic logic. This approach enhances pharmacovigilance by reducing noise and improving reasoning accuracy in clinical data.
Area of Science:
- Biomedical Informatics
- Computational Linguistics
- Pharmacovigilance
Background:
- Automated drug-drug interaction (DDI) extraction is crucial for pharmacovigilance.
- Current models struggle with the linguistic paradox of localized cues and long-range semantic logic in DDIs.
- Transformer models face challenges in disentangling local descriptors from distal logical chains, leading to errors.
Purpose of the Study:
- To present DuSSM, a parallel state-space framework designed to structurally disentangle surface patterns from semantic evolution for DDI extraction.
- To address the limitations of monolithic architectures in resolving feature entanglement in clinical text.
- To improve the accuracy and efficiency of automated DDI extraction.
Main Methods:
- DuSSM employs a bifurcated pipeline with an explicit convolutional stream for local pattern recognition and an implicit stream using selective state-space modeling (Mamba) for semantic stability.
- The framework decouples downstream reasoning, achieving linear ๐ช(N) complexity after initial quadratic ๐ช(N2) encoding.
- Experiments were conducted on four biomedical benchmarks: DDI-2013, ChemProt, GAD, and EU-ADR.
Main Results:
- DuSSM achieved a robust F1-score of 82.27% on the DDI benchmark, demonstrating strong generalization across diverse datasets.
- The model exhibited high precision (94.32%) in identifying non-interaction cases, effectively reducing alert fatigue.
- The dual-stream approach successfully reconciled computational efficiency with mechanistic interpretability.
Conclusions:
- DuSSM offers a scalable and trustworthy paradigm for deciphering complex biological interactions in electronic health records.
- The framework's ability to disentangle linguistic features enhances the reliability of automated pharmacovigilance systems.
- DuSSM provides a significant advancement in handling the complexities of DDI extraction, improving clinical decision support.
