Large Language Models (LLMs) exhibit considerable potential in medical applications; however, their susceptibility to subtle linguistic variations—such as differences in terminology, colloquial phrasing, and word order—poses a significant challenge to clinical reliability. To address this, we introduce MeDRNet, a knowledge-enhanced, multi-model medical AI framework that dynamically integrates general-purpose and domain-specific models through an adaptive routing mechanism. Its modular architecture adopts a sandwich-style knowledge enhancement, integrating medical knowledge graphs, retrieval-augmented generation, and hidden-layer fusion to strengthen factual grounding. On top of this layered design, adversarial training, logical consistency constraints, and a knowledge alignment module are further employed to reduce hallucinations and enhance robustness. Designed to maintain semantic stability across diverse clinical inputs, including noisy, informal, and domain-specific queries, MeDRNet is well-suited for high-stakes healthcare scenarios. Comprehensive experiments on PromptCBLUE, MultiMedBench, and a real-world query set demonstrate that MeDRNet consistently surpasses robust baselines—such as GPT-4, Aquila-Med LLM, and HuatuoGPT-o1—in accuracy, robustness, and hallucination resistance. These findings position MeDRNet as a scalable and reliable foundation for clinical language understanding tasks. Furthermore, the framework is readily adaptable to downstream applications, including diagnostic decision support, electronic health record (EHR) summarization, and multilingual medical question answering, paving the way for seamless integration of LLMs into next-generation clinical workflows.