When a doctor faces complex symptoms, every additional diagnostic test has a cost, both financial and in time. Yet Large Language Models, despite their vast medical knowledge, tend to over-order tests because they struggle to reason systematically under cost constraints. GraphDx, a new multi-agent framework, aims to bridge this “knowledge-reasoning gap” with an unconventional architecture.

GraphDx introduces two core innovations. The first is an automated pipeline that uses LLMs to build Medical Diagnosis Knowledge Graphs (MDKGs): graphs with quantized typicality, action-centric topology, and dual-objective attributes — diagnostic relevance plus cost sensitivity. The second is a trio of collaborative agents: Perception, Reasoning, and Decision. Perception and Decision handle language understanding and generation, while Reasoning performs deterministic evidence scoring and cost-aware planning on the graph. In effect, the “cognitive” part is separated from the decision logic, reducing hallucinations and unnecessary tests.

Tests on MedQA and MIMIC-IV, run with three different LLM backbones — DeepSeek-V3, Kimi-k2, and Llama-3.3 — show clear results: diagnostic success rates jump from 50-68% to 79-93%, while test costs fall by 20-54%. Those numbers signal more than incremental improvement; they shift the problem’s structure itself.

For those working on deployments with sensitive data, like hospitals, the most relevant detail is that the deterministic reasoning on the graph does not require hundreds of billions of parameters for every inference step. The heavy computational work is absorbed once by the graph-building pipeline; the Reasoning agent then operates on symbolic structures, easing the dependence on continuous LLM inference. This lowers the hardware bar and paves the way for on-premise, local installations, without needing to send patient data to the cloud. With open, consumer-grade models like those used, the framework already runs on manageable infrastructure — a sign that supercomputers are not needed.

Economically, the reduction in ordered tests translates, in a cloud scenario, to fewer API calls. But in a self-hosted context, the savings are structural: fewer inference requests mean less GPU and energy consumption, with a direct impact on Total Cost of Ownership. GraphDx shows that a hybrid approach — LLMs for language and hypotheses, a graph for planning — can be more efficient than a purely generative pipeline, and with decision transparency that in medicine is not optional.

Of course, the leap from dataset benchmarks to real-world clinical practice is substantial and will require further validation. But the signal is strong: diagnostic automation does not have to choose between accuracy and sustainability. With agentic architectures and knowledge graphs, a path can be built toward more interpretable, economical, and privacy-conscious systems. For those designing digital healthcare infrastructure, it is a shift in perspective worth watching.