A team led by Shaoting Tan has introduced GraphDx, a knowledge-enhanced multi-agent framework that tackles one of the most stubborn problems in medical AI: getting large language models to think like cost-conscious clinicians. In experiments across MedQA and MIMIC-IV, the system raised diagnostic success rates from a baseline of 50-68% all the way to 79-93%, while simultaneously cutting the cost of diagnostic tests by 20-54%.
Why LLMs Struggle With Real-World Diagnosis
Large language models have swallowed enormous volumes of medical literature, yet they still fumble when asked to order tests the way a seasoned doctor would. The issue is not a lack of knowledge. It is a gap between knowing facts and reasoning under pressure. Current LLM approaches tend to over-test, treating every symptom as a prompt for another expensive lab or scan rather than weighing whether the information is worth the price. In real hospitals, that behavior burns budgets and strains already thin resources.
How GraphDx Closes the Gap
GraphDx attacks the problem with two linked innovations. The first is an automated pipeline that turns raw medical knowledge into structured Medical Diagnosis Knowledge Graphs (MDKGs). These graphs are not simple concept maps. They encode quantized typicality (how common a symptom is for a given disease), an action-centric topology (which tests reveal which evidence), and dual-objective attributes that track both diagnostic relevance and cost. In other words, the graph knows not only that a chest X-ray helps rule out pneumonia, but also how much that X-ray costs and how strongly the result would shift the diagnosis.
The second innovation is a trio of collaborative agents: Perception, Reasoning, and Decision. The Perception and Decision agents handle the messy work of understanding patient descriptions and generating human-readable output. The Reasoning agent does the heavy lifting on the graph itself, running deterministic evidence scoring and cost-aware planning. By separating probabilistic language tasks from deterministic graph operations, GraphDx avoids the hallucination and cost-blindness that plague end-to-end LLM diagnosis.
Benchmark Results Across Three Major Models
The researchers tested GraphDx on top of three distinct LLM backbones: DeepSeek-V3, Kimi-k2, and Llama-3.3. The gains were consistent across all three, which suggests the framework is model-agnostic rather than dependent on a single proprietary system. That portability matters. Hospitals and health-tech startups are wary of locking into one vendor, and a method that works on open-weight models like Llama-3.3 lowers the barrier to real deployment.
The cost reductions are arguably the bigger headline. A 54% drop in test spending, paired with higher accuracy, flips the usual AI healthcare narrative. Too often, new tools add expense by layering software on top of existing workflows. GraphDx shows that smarter reasoning can actually shrink the bill.
What This Means for Clinical AI
The framework arrives at a moment when healthcare systems worldwide are searching for ways to control spending without sacrificing outcomes. Diagnostic testing accounts for a significant slice of hospital budgets, and unnecessary tests also expose patients to avoidable risks, from radiation to false positives that trigger cascades of further procedures. A system that reasons explicitly about cost, rather than treating it as an afterthought, could reshape how AI is adopted in clinical settings.
There is also a transparency angle. Because the Reasoning agent operates on an explicit knowledge graph, clinicians can trace why GraphDx recommended a particular test. That interpretability is a regulatory and trust advantage over black-box models that simply output a diagnosis and a test order.
The Road Ahead
The team has published the work on arXiv, and the next steps will likely involve real-world pilots in controlled clinical environments. MedQA and MIMIC-IV are rigorous benchmarks, but live patient data introduces noise that no benchmark fully captures. If GraphDx holds up outside the lab, it could become a template for how multi-agent architectures handle other resource-constrained decisions in healthcare, from triage to treatment planning. Watch for follow-up studies that test the framework against diverse patient populations and cost structures.