MedGuard: An AI Gatekeeper Embedding Fact-Checking into Clinical Workflows

In online consultations, a statement like 'Your fatty liver is nothing to worry about' may hide serious risks. The report mentions a patient with a history of diffuse large B-cell lymphoma, where the doctor misinterpreted fatty liver features as 'high density on imaging' based on the imaging description, thereby ruling out fatty liver, but the key medical facts were inconsistent with the actual imaging findings. Traditional safety solutions rely on keyword rules or simple text classification, which are effective for 'obvious errors' but struggle with edge cases that require integrating the patient's global context, clinical guidelines, and evidence chains.

MedGuard is jointly proposed by Ant AI Safety Lab and Xiamen University, and the paper has been accepted by npj Digital Medicine. It is not another AI consultation system but a layer of safety infrastructure that can be embedded into medical service workflows, verifying key medical facts at the evidence level before doctors provide diagnostic opinions, AI generates responses, or medication plans are formed.

The research categorizes medical fact-checking in online diagnosis and treatment into three core challenges: key information being overwhelmed by long conversations, difficulty in defining risk boundaries, and the need for traceability of medical judgment results. MedGuard's approach is to decompose complex doctor-patient communication into independently verifiable atomic medical claims, maintain a global patient context, extract key information such as age, gender, medical history, allergy history, and current medications, and rewrite statements scattered across multiple rounds of dialogue by de-contextualizing them, completing references and omitted information while preserving critical clinical constraints like negations, dosages, and timing.

MedGuard outputs three states for each medical claim: No Risk, Low Risk, and High Risk, where Low Risk indicates insufficient evidence or semantic ambiguity requiring further verification. High-confidence No Risk and High Risk claims follow a fast path for initial judgment, and only claims entering the uncertain region trigger external medical evidence retrieval. This uncertainty-driven verification mechanism focuses limited resources on the most critical edge cases, reducing missed detections while controlling false alarms.