by Yaxun Jia, Zhu Yuan, Lian Zhu, Bing Han, Li Ren, Zuo-lin XiangAutomated drug-drug interaction (DDI) extraction is a cornerstone of global pharmacovigilance, yet its progress is stymied by a fundamental linguistic paradox: relations are signaled by localized morphological cues while being governed by long-range semantic logic. Current monolithic architectures, including Transformer-based models, often face challenges in resolving this feature entanglement, where local clinical descriptors often distort the distal logical chain, leading to noise propagation and reasoning failures. To address this, we present DuSSM, a parallel state-space framework that structurally disentangles surface patterns from semantic evolution. DuSSM implements a bifurcated pipeline: an explicit convolutional stream acting as a local pattern recognizer to isolate syntactic triggers, and an implicit stream leveraging selective state-space modeling (Mamba) to maintain stable semantic states. Although the initial contextual encoding retains a quadratic complexity (𝒪(N2)), this decoupled downstream reasoning module operates with a strictly linear 𝒪(N) complexity. Our extensive experiments across four diverse biomedical benchmarks (DDI-2013, ChemProt, GAD, EU-ADR) demonstrate that this dual-stream feature separation generalizes exceptionally well, achieving a robust Fl-score of 82.27% on the DDI benchmark. Notably, DuSSM yields 94.32% precision on non-interaction cases, effectively mitigating the alert fatigue that can arise from the probabilistic nature of generative large language models (LLMs) in zero-shot settings. By reconciling computational efficiency with mechanistic interpretability, DuSSM provides a scalable and trustworthy paradigm for deciphering complex biological interactions within massive electronic health records. All code and data have been publicly released at: https://github.com/Hero-Legend/DuSSM.