MedGuard: an LLM-based gatekeeper for detecting clinical risks in Chinese telemedicine consultations

Telemedicine embodies both promise and peril, offering unprecedented convenience while exposing patients and practitioners to potential safety risks. To address these concerns, we present MedGuard, the first LLM-based agent to monitor telemedicine consultations and proactively flag risks in diagnoses, prescriptions, and other clinical dialogs. MedGuard makes three key contributions: (1) a dedicated workflow specifically designed for telemedicine risk alerts; (2) a comprehensive benchmark covering real and synthesized medical consultations with fine-grained annotations, together with a structured medical knowledge base; and (3) methodological innovations, including global-memory–driven claim extraction, self-evolved retriever, cautious hierarchical classifier, training data generation for domain-transfer reinforcement learning. They could be beneficial for solving the data scarcity problem in the medical domain. Extensive benchmark experiments on both real‑world and synthetic datasets demonstrate that MedGuard achieves, on average, 21.20–23.02% higher performance than state‑of‑the‑art agentic systems across multiple evaluation metrics. Human evaluations involving over one hundred medical professionals further confirm that MedGuard enhances telemedicine efficiency, reduces diagnostic errors, and facilitates online communication, underscoring its potential for integration into real‑world telemedicine workflows.