💻 computer science

Sparse Multi-Stage Expert-Agent Routing for Complex Clinical Reasoning

This paper proposes Sparse Multi-Stage Expert-Agent Routing, a framework that adaptively activates a minimal set of medical expert agents across evolving clinical stages to achieve high diagnostic accuracy with significantly reduced resource usage, alongside a new fact-aware evaluation metric (ClinFEScore) that strongly correlates with clinician judgments.

Original authors: Sike Xiang, Shuang Chen, Qian sun, Jia Cheng, Yusi Wei, Amir Atapour-Abarghouei

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sike Xiang, Shuang Chen, Qian sun, Jia Cheng, Yusi Wei, Amir Atapour-Abarghouei

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Diagnosing a complex illness is rarely a single moment of clarity. It is a slow, unfolding story where a patient's symptoms, blood tests, imaging scans, and follow-up observations arrive one by one, often over days or weeks. As new facts emerge, doctors must constantly revise their thinking, sometimes bringing in specialists from different fields to help interpret the changing picture. This process, known as clinical reasoning, is difficult to automate because it requires a system that can learn from new information as it arrives, rather than just guessing the answer from a static list of symptoms. Current artificial intelligence tools often struggle with this; they tend to either make a single guess based on the first piece of information they see, or they call upon a fixed team of virtual experts to discuss every case from start to finish, regardless of whether all those experts are actually needed. This approach is inefficient, wasting time and computational power, and it fails to mimic the natural, evolving way human doctors solve difficult medical puzzles.

Researchers at Durham University have developed a new method to teach artificial intelligence how to reason through complex medical cases more like a human team does. They created a system called Sparse Multi-Stage Expert-Agent Routing. Instead of asking a large group of virtual specialists to review every case in its entirety, this system breaks the diagnosis down into stages, matching the way medical evidence actually becomes available. At each stage, the system looks at the current state of the patient's case and decides which specific expert is needed right now. It might start with a general practitioner to assess initial symptoms, then call in a radiologist only when an X-ray becomes available, and perhaps a geneticist later if a specific test result points that way. Crucially, the system is designed to be "sparse," meaning it activates only the few experts who are truly necessary at that moment, rather than exhausting every possible resource.

To test this idea, the researchers built a framework where the artificial intelligence maintains a running record of the case, updating its understanding as new evidence is added. They trained the system on reconstructed medical cases where information was presented in a sequence, mimicking a real hospital investigation. The system learned to assign an "activation score" to each potential expert agent. If the score was high enough, that expert would be invited to contribute their reasoning; if not, they remained silent. This decision-making process was supported by a memory system that allowed experts to remember their previous contributions without needing to be re-consulted unnecessarily. The team also invented a new way to measure the quality of the AI's final diagnosis. Instead of just checking if the AI's words matched the correct answer, they developed a tool called ClinFEScore that checks whether the AI's conclusion was supported by the actual medical facts presented in the case, ensuring the reasoning was sound and not just a lucky guess.

The results of their experiments were striking. When compared to a system that called upon all available experts at every stage, the new sparse routing method reduced the average number of experts involved in a single case from 17 down to just 3. Despite using far fewer resources, the quality of the diagnosis remained high, with the system maintaining strong accuracy in identifying the correct medical facts. In tests using simulated data, the new method lowered the estimated energy cost of the computation significantly while keeping the diagnostic quality comparable to the much more expensive, fully staffed approach. The researchers also tested their system on 200 real-world cases from hospital multidisciplinary teams. In these real-life scenarios, the system achieved a diagnostic accuracy of 91.5%, verified by human doctors. Furthermore, the new evaluation tool they created, ClinFEScore, showed a very strong agreement with how human doctors judged the quality of the reasoning, confirming that the system was not just producing text that looked right, but was actually reasoning correctly based on the evidence.

The study suggests that the key to efficient medical artificial intelligence is not simply having more experts or more computing power, but having a smarter way to coordinate them. By making the participation of experts conditional on the evidence available at each step, the system avoids the redundancy of asking the same questions repeatedly or consulting specialists who have nothing to add. The researchers found that this approach worked well across different types of underlying AI models, indicating that the improvement came from the routing strategy itself rather than the specific language model used. While the work is a significant step forward, the authors note that future research will need to test this method across a wider variety of hospitals and medical workflows to ensure it remains reliable in diverse real-world settings. For now, the findings offer a promising path toward AI systems that can handle complex medical reasoning with the same efficiency and adaptability as a skilled human team.

Technical Summary: Sparse Multi-Stage Expert-Agent Routing for Complex Clinical Reasoning

Problem Statement

Complex clinical diagnosis is inherently a dynamic, multi-stage process where evidence accumulates over time (e.g., initial symptoms, lab results, imaging, pathology, follow-ups). Clinicians continuously revise diagnostic hypotheses as new data arrives and often consult different medical specialities depending on the evolving nature of the case. However, existing Large Language Model (LLM)-based clinical reasoning systems typically operate in one of two suboptimal modes:

Single-pass prediction: Processing the entire case description at once, failing to model the temporal evolution of evidence.
Fixed multi-agent workflows: Engaging a static set of agents regardless of the specific evidence stage, leading to either static expert participation or unnecessarily exhaustive consultation.

These approaches fail to couple expert participation with the gradual arrival of clinical evidence, resulting in inefficient resource usage, redundant opinions, and a lack of adaptability to changing diagnostic questions. Furthermore, evaluating the output of such systems is challenging; standard metrics often rely on surface-level text similarity rather than factual consistency with clinical evidence.

Methodology

The authors propose Sparse Multi-Stage Expert-Agent Routing, a framework that models diagnosis as a stage-wise routing process where expert activation is conditional on the current case state and available evidence.

1. Framework Architecture

The system processes multimodal clinical inputs (demographics, symptoms, labs, imaging, etc.) organized into sequential evidence blocks.

Stage-wise State Evolution: The system maintains an evolving case representation $h^{(s)}$ . At each stage $s$ , new evidence $x^{(s)}$ is encoded and concatenated with the previous state to form a cross-stage representation.
Dynamic Routing Mechanism: A routing module assigns an activation score to each candidate expert agent based on the current state.
- Internal State Update: Routing states are updated over internal steps $t$ using a decay coefficient $\lambda$ , allowing activation to emerge progressively rather than via static one-shot gating.
- Adaptive Thresholding: To prevent persistent over-activation, an adaptive threshold $\theta$ is applied, which increases based on recent activation history ( $a^{(s)}$ ).
- Population-Level Inhibition: A mechanism controls the mean activation rate $r^{(s)}$ across all agents. If the population activity exceeds a target $\rho$ , a lateral inhibition term $\eta$ raises the selection threshold, ensuring sparsity.
- Binary Masking: A binary mask $m^{(s)}_i$ is derived from the continuous scores to determine which specific agents are invoked for reasoning at that stage.
Expert Memory: The framework utilizes expert-specific memory to preserve professional roles and relevant case context across stages, allowing later reasoning to build on earlier findings without re-consulting the full pool.

2. Training Objective

The model is trained using a joint loss function $\mathcal{L}$ that balances diagnostic quality with routing efficiency:

Diagnostic Consistency ( $L_{final}$ ): Measures semantic consistency between the final output and the reference.
Energy Cost ( $L_{energy}$ ): Penalizes the invocation cost of experts and the overhead of activating any agent in a stage.
Entropy Regularization ( $L_{entropy}$ ): Encourages diverse activation patterns to avoid mode collapse.
Activation Rate Constraint ( $L_{rate}$ ): Enforces a target average activation rate to maintain sparsity.

3. Evaluation Protocol: ClinFEScore

To address the limitations of surface-level similarity metrics, the authors introduce ClinFEScore, a fact-aware semantic evaluation protocol.

Process: An LLM-based extractor identifies clinically important factual statements from both the generated conclusion and the reference answer.
Metrics: These facts are encoded into a shared biomedical semantic space to calculate:
- Fact Similarity: Overall semantic overlap.
- Fact Precision: How well the reference supports the generated facts.
- Fact Recall: How well the generated conclusion covers the reference facts.

Key Contributions

Sparse Multi-Stage Expert-Agent Routing: A stage-aware framework where expert activation is an evidence-dependent decision updated throughout the case, rather than a fixed workflow.
ClinFEScore: A novel evaluation protocol for free-text clinical conclusions that measures factual consistency and coverage, validated against clinician judgments.
Empirical Validation: Demonstrations that sparse routing preserves diagnostic quality while significantly reducing computational and consultation costs, validated on both reconstructed datasets and real-world hospital cases.

Experimental Results

The framework was evaluated on reconstructed multi-stage cases from MAC and AgentClinic-NEJM, as well as 200 real-world hospital Multidisciplinary Team (MDT) cases.

Quantitative Performance (MAC & AgentClinic-NEJM)

Efficiency: The sparse routing framework reduced the average number of activated expert agents from 17.0 (full multi-stage consultation) to 3.0, while lowering the simulated energy cost from 34.2 to 7.15.
Quality: Despite the reduction in agents, the framework maintained or improved diagnostic quality. The best configuration (Sparse + Memory) achieved ClinFEScore sub-metrics of 0.9940 (Similarity), 0.9792 (Precision), and 0.9780 (Recall), outperforming full consultation models (0.9845, 0.9706, 0.9690) and single-agent baselines.
Ablation: The framework proved robust across different backbone models (including MedGemma-4B and Hulu-Med-4B), suggesting the gains are driven by the routing architecture rather than a specific LLM. Deeper routing layers (4 vs. 2) did not yield significant gains and led to pattern collapse.

Real-World MDT Evaluation

Clinician Agreement: Three clinicians assessed the generated conclusions on 200 real cases, showing strong agreement (mean pairwise weighted $\kappa = 0.93$ ).
Metric Correlation: ClinFEScore correlated strongly with clinician judgments (Spearman's $\rho = 0.81$ ; Pearson's $r = 0.87$ ).
Accuracy & Cost: The framework achieved 91.5% clinician-verified diagnostic accuracy with approximately five expert-agent/LLM calls per case. This outperformed MedAgents (166/200 correct, ~17 calls) and MDAgents (174/200 correct, ~7 calls).

Significance and Claims

The paper claims that sparse stage-wise coordination is a clinically relevant and efficient approach to LLM-based clinical reasoning. By treating expert participation as a dynamic decision linked to evidence progression, the system avoids the computational and clinical inefficiencies of exhaustive consultation. The introduction of ClinFEScore provides a more rigorous standard for evaluating clinical reasoning outputs, moving beyond superficial text matching to factual consistency. The authors conclude that their approach successfully balances diagnostic quality with resource constraints, making it a viable model for real-world clinical decision support where expert availability is limited.