GraphDx: A Cost-Aware Knowledge-Enhanced Multi-Agent Framework for Sequential Diagnosis
Original reporting by arXiv (cs.AI)
Sequential diagnosis is a critical process in medicine that involves iteratively gathering information to refine a diagnosis, balancing diagnostic accuracy against the imperative of managing resource costs. While Large Language Models (LLMs) encode vast amounts of medical knowledge, they have historically struggled with this nuanced challenge, often exhibiting a "knowledge-reasoning gap" that leads to excessive and costly testing without systematic, cost-aware planning.
GraphDx's Approach
A new framework, GraphDx, addresses this persistent challenge through two core innovations. First, it introduces an automated pipeline that leverages LLMs to construct Medical Diagnosis Knowledge Graphs (MDKGs). These MDKGs are uniquely designed with quantized typicality, an action-centric topology, and dual-objective attributes that factor in both diagnostic relevance and cost-sensitivity. Second, GraphDx employs three collaborative agents—Perception, Reasoning, and Decision—where the Reasoning Agent deterministically scores evidence and performs cost-aware planning directly on the MDKG. Tested across established LLM backbones like DeepSeek-V3, Kimi-k2, and Llama-3.3 on MedQA and MIMIC-IV datasets, GraphDx dramatically improved diagnostic success rates from 50–68% to an impressive 79–93%, concurrently reducing test costs by 20–54%. This development signals a significant step towards robust, economical, and interpretable automated clinical diagnosis.
The GraphDx framework represents a significant advance in AI-driven sequential diagnosis, directly addressing the critical knowledge-reasoning gap observed in traditional large language models. By integrating sophisticated Medical Diagnosis Knowledge Graphs (MDKGs) with a multi-agent system, GraphDx moves beyond the limitations of relying solely on pattern recognition, enabling truly systematic and cost-aware diagnostic pathways. Its demonstrated ability to boost diagnostic success rates to over 90% while simultaneously cutting test costs by up to 54% underscores its potential to deliver both highly accurate and economically viable solutions. This research not only provides a robust and interpretable automated clinical diagnosis tool but also establishes a new benchmark for balancing performance with practical resource considerations.
Reshaping Clinical AI The implications of GraphDx extend far beyond its immediate impressive metrics. Its success in embedding cost-aware reasoning directly into the diagnostic process heralds a future where AI systems can be seamlessly integrated into resource-constrained clinical environments. This shift from knowledge-rich but reasoning-poor LLMs to architecturally sound, interpretable decision-making frameworks could fundamentally reshape how physicians leverage AI, moving towards collaborative tools that suggest optimal, evidence-based, and economical diagnostic sequences rather than merely generating probabilities. Such advancements promise more efficient healthcare delivery, reduced burdens on both patients and providers, and ultimately, a path to more accessible and effective medical care globally. GraphDx marks a pivotal step toward AI solutions that are not just intelligent, but also practical, ethical, and deeply aligned with the realities of clinical practice.
Frequently asked questions
- What problem does GraphDx address in automated medical diagnosis using large language models?
- GraphDx addresses the critical challenge of sequential medical diagnosis where Large Language Models (LLMs) struggle to balance diagnostic accuracy with resource costs. Existing LLMs often exhibit a knowledge-reasoning gap, leading to excessive or redundant testing. GraphDx aims to provide a more systematic and cost-effective approach by integrating structured medical knowledge and intelligent planning to guide the diagnostic process efficiently and reduce unnecessary expenses.
- How does the GraphDx framework enhance the diagnostic capabilities of AI in healthcare?
- GraphDx enhances AI diagnostic capabilities through two core innovations: Medical Diagnosis Knowledge Graphs (MDKGs) and collaborative agents. MDKGs provide structured, cost-aware medical knowledge, enabling the system to understand diagnostic relevance and cost-sensitivity. A dedicated Reasoning Agent then uses this graph for deterministic evidence scoring and cost-aware planning, allowing AI to reason systematically, avoid redundant tests, and make more accurate and resource-efficient diagnoses in clinical settings.
- What are the primary benefits of using GraphDx for sequential clinical diagnosis?
- GraphDx offers significant benefits for sequential clinical diagnosis, primarily by improving diagnostic success rates while substantially reducing test costs. Experiments demonstrate it can increase diagnostic accuracy from 50-68% to 79-93% and cut test expenses by 20-54%. This provides a more robust, economical, and interpretable solution, ensuring that AI-powered diagnosis is both highly effective in identifying conditions and efficient in resource utilization within healthcare.