Source-linked AI summary
Sparse Multi-Stage Expert-Agent Routing for Complex Clinical Reasoning
Sike Xiang, Shuang Chen, Qian sun, Jia Cheng, Yusi Wei, Amir Atapour-Abarghouei
TL;DR
Complex clinical reasoning requires updating diagnoses as evidence accumulates and coordinating specialists without static or exhaustive consultation. The paper introduces stage-wise sparse expert routing with evolving case state, expert memory, and fact-aware evaluation. It reports fewer activated experts while preserving diagnostic quality, with strong agreement between ClinFEScore and clinician judgements on real-world cases.
Problem
Existing LLM clinical reasoning systems commonly use single-pass prediction or fixed multi-agent workflows, limiting adaptive expert participation as evidence evolves.
Method
Sparse Multi-Stage Expert-Agent Routing updates an evolving case state and adaptively activates expert agents from progressively available multimodal evidence, while ClinFEScore evaluates factual semantic quality.
Results
17.0 to approximately 3.0 activated experts: the framework reduces average expert invocation while maintaining diagnostic quality, and reaches 91.5% clinician-verified diagnostic accuracy on 200 MDT cases.
Takeaways & Limitations
The findings support sparse stage-wise routing as a way to retain clinical content while avoiding unnecessary expert participation.
Takeaways & Limitations
Future validation must test broader cross-institutional and prospective settings, longitudinal evidence, workflow variation, and additional clinical safeguards.
Abstract
from arXiv · showhide
Complex clinical reasoning requires models to update diagnostic hypotheses as new evidence emerges and to coordinate different medical specialities under limited consultation resources. Existing LLM-based clinical reasoning systems typically perform single-pass prediction or rely on fixed multi-agent workflows, making expert participation either static or unnecessarily exhaustive. We propose Sparse Multi-Stage Expert-Agent Routing, a language-based clinical reasoning framework that models diagnosis as a stage-wise routing process. Given progressively available clinical evidence derived from multiple modalities, the framework maintains an evolving case state and adaptively activates a sparse set of medical expert agents, supported by expert-specific memory across stages. To evaluate free-text diagnostic conclusions beyond surface similarity, we further introduce ClinFEScore, a fact-aware semantic evaluation protocol for clinical reasoning outputs. On reconstructed multi-stage cases from MAC and AgentClinic-NEJM, our framework reduces the average number of activated experts from 17.0 to 3.0 whilst maintaining strong fact-level diagnostic quality. On 200 real-world hospital MDT cases, ClinFEScore correlates strongly with clinician judgements (Spearman's $ρ=0.81$; Pearson's $r=0.87$), whilst our method achieves 91.5\% clinician-verified diagnostic accuracy with approximately five expert-agent/LLM calls per case. These results support sparse stage-wise coordination as an efficient and clinically relevant approach to LLM-based clinical reasoning.
1 Introduction
Clinical diagnosis unfolds as evidence accumulates, but many LLM systems use single-pass or fixed expert participation. The proposed framework routes a sparse, adaptive set of experts across stages while evaluating factual diagnostic quality.
- Motivation: Clinical evidence arrives progressively, requiring diagnostic hypotheses to be revised and specialist participation to change as the case becomes more specific.The motivating evidence includes symptoms and examination first, followed by laboratory, imaging, pathology, and follow-up information.
- Motivation: Invoking every expert at every stage is costly, redundant, and potentially exposes agents to information irrelevant to their roles.The framework therefore treats consultation as conditional on the evidence available and prior inferences.
- Framework: Sparse Multi-Stage Expert-Agent Routing updates expert activation using the evolving case state, multimodal evidence, adaptive thresholds, and population-level inhibition.Activated experts produce stage-wise reasoning, whose outputs update the case state; expert-specific memory preserves relevant context across stages.
- Results: 17.0 to 3.0 activated experts: the complete framework reduces average consultation while maintaining comparable or better ClinFEScore results.The evaluation uses sequential evidence blocks reconstructed from MAC and AgentClinic-NEJM, with later information excluded from earlier stages.
- Results: 91.5% clinician-verified diagnostic accuracy is achieved with approximately five expert-agent/LLM calls per case.ClinFEScore correlates with clinician judgements at Spearman’s ρ = 0.81 and Pearson’s r = 0.87.
- Evaluation: ClinFEScore evaluates the semantic consistency and completeness of free-text clinical conclusions at the factual level and is validated against clinician judgements.Its agreement is assessed on 200 real-world multidisciplinary team cases.
2 Related Work
Related work spans complex clinical reasoning, medical multi-agent systems, and conditional routing. The paper positions its approach against methods that handle only parts of the evolving, collaborative diagnostic process.
- Complex clinical reasoning: Complex clinical reasoning integrates heterogeneous evidence and revises diagnostic judgement as information becomes available.It can also involve multiple specialities jointly interpreting evolving evidence.
- Medical multi-agent reasoning: Recent clinical multi-agent systems use specialised roles for diagnosis, simulated consultation, and broader collaborative reasoning.Role decomposition can improve robustness and reduce some limitations of single-model inference.
- Routing mechanisms: Mixture-of-experts and routing methods provide conditional computation, while medical multi-agent systems often organise reasoning around specialist roles or predefined consultation processes.The paper identifies these approaches as related components rather than complete solutions to stage-wise clinical routing.
3 Method
The framework organizes multimodal clinical evidence into stages, maintains cross-stage case state, and selectively activates expert agents before aggregating their outputs into a final diagnosis.
- Multimodal clinical inputs are organized into structured stage-wise evidence representations for sequential reasoning.
- At each stage, the router uses the current case state and evidence to selectively activate relevant medical expert agents.
- Routing states are updated across internal steps with decayed historical information, enabling progressively emerging expert-agent activation rather than static one-shot gating.
- Adaptive thresholds and population-level inhibition produce a binary activation mask that determines which experts are invoked for stage-wise reasoning.
- Later stages combine prior intermediate state with newly encoded evidence, and stage outputs are aggregated into the final diagnostic conclusion.
- Training jointly optimizes final diagnostic consistency with invocation energy, activation entropy, and average activation-rate regularization.
4 Experiments
Experiments reconstruct clinical cases as sequential evidence blocks and evaluate both diagnostic quality and routing efficiency using ClinFEScore across system configurations.
- Cases from MAC and AgentClinic-NEJM are reconstructed into sequential stage-wise inputs reflecting the order in which clinical information becomes available.
- The framework uses a fixed expert-agent pool and trains the routing module on an 8:2 training-validation split with 20 epochs and four internal routing steps per stage.
- ClinFEScore evaluates diagnostic quality with fact-aware semantic measures adapted to complex multi-speciality clinical reasoning.
- 0.9972, 0.9904 and 0.9887 are the Ground Truth values for Fact Similarity, Fact Precision and Fact Recall, respectively, serving as an upper-bound reference.
- 3.0119 activated expert agents and 7.1488 energy cost are achieved by sparse routing without memory augmentation, with scores of 0.9936, 0.9788 and 0.9768.
4.5 Ablation Studies
The ablations examine backbone choice, routing depth, and real-world clinical performance under controlled evaluation settings. Results indicate that sparse routing remains effective across backbones, while deeper routing adds cost without substantive gains.
- Backbone Ablation: Table 2 compares general-purpose and medical-domain backbones under matched routing and memory configurations.Medical-domain models are deployed locally, so retrieval-dependent expert memory is unavailable, while short-term and role-specific memory remain active.
- Backbone Ablation: MedGemma-4B achieves the best overall results at 0.9941, 0.9796 and 0.9786 despite lacking retrieval-based expert memory.The performance gap among mainstream backbone models is relatively small, while domain alignment keeps the smaller medical models competitive.
- Routing Layers Ablation: Four-layer routing produces only marginally different ClinFEScore values from two-layer routing: 0.9940, 0.9792 and 0.9780 versus 0.9941, 0.9798 and 0.9785.The four-layer configuration is not naturally aligned with the dataset’s two-stage reasoning process.
- Routing Layers Ablation: Four-layer routing increases activated experts and energy cost to 4.0000 and 9.4000, compared with 3.0119 and 7.1488 for two-layer routing.Activation dynamics also collapse toward general_internal in early layers and radiology in later layers.
- Real-World MDT Evaluation: On 200 real-world MDT cases, the method correctly diagnoses 183 cases, achieving 91.5% accuracy with approximately five calls per case.MDAgents achieves 174 cases with approximately seven calls, while MedAgents achieves 166 cases with approximately seventeen calls.
5 Conclusion
The paper concludes that sparse multi-stage expert-agent routing combines staged clinical information, selective activation, collaborative reasoning, and fact-aware evaluation. It reports reduced consultation demand and strong real-world diagnostic performance, while identifying broader validation as future work.
- Conclusion: The framework combines stage-wise clinical information organisation, sparse expert-agent activation, collaborative reasoning, and fact-aware semantic evaluation.ClinFEScore assesses semantic consistency and completeness of free-text clinical conclusions at the factual level.
- Conclusion: The average number of activated expert agents falls from 17.0 to approximately 3.0 while diagnostic quality is maintained.The conclusion attributes complementary benefits to multi-stage modelling, sparse routing, and expert memory.
- Conclusion: On real-world MDT cases, the framework achieves 91.5% clinician-verified diagnostic accuracy with approximately five expert-agent or LLM calls per case.ClinFEScore also correlates strongly with clinician factual-consistency judgements.
- Conclusion: Future work targets broader cross-institutional and prospective validation across longitudinal evidence, speciality activation patterns, and clinical workflows.The authors also plan stricter factual-consistency constraints and more detailed token, latency, and financial cost measurements.
A Domain Skill Design
Domain skills are implemented as structured role-specific prompts rather than learned through additional training. These prompts stabilise each expert agent’s professional perspective and reasoning scope.
- A Domain Skill Design: Each expert agent receives skill descriptions aligned with its professional identity, responsibility scope, clinical focus, evidence preferences, and reasoning style.The skills are instantiated as structured prompt templates and supplied to the backbone model.
- A Domain Skill Design: Role-specific professional prompts provide stable prior knowledge for each expert agent without additional training.This design supports professional specialisation and role consistency during multi-expert collaboration.
- A Domain Skill Design: Figure 3 depicts memory as a triangular interaction among expert memory, short-term memory, and long-term memory across reasoning stages.The figure presents the memory structure as supporting stable yet adaptive reasoning.
B Expert Memory Design
The expert-memory module combines stable domain priors, evolving case context, and reusable cross-case experience. Short-term memory is updated during reasoning, while long-term memory consolidates and retrieves prior patterns.
- B Expert Memory Design: The memory module forms a triangular interaction among expert memory, short-term memory, and long-term memory.This structure avoids treating each expert invocation as isolated from prior reasoning and historical experience.
- B Expert Memory Design: Expert memory anchors each agent to its domain role, reasoning scope, preferred evidence sources, and high-priority diagnostic cues.It helps preserve speciality alignment as clinical evidence changes across stages.
- B Expert Memory Design: Short-term memory stores extracted evidence, intermediate judgements, salient features, and exchanged expert outputs for the current case.It is updated dynamically so later stages build on earlier reasoning rather than restarting.
- B Expert Memory Design: Long-term memory consolidates valuable information from current cases into reusable cross-case knowledge and retrieves patterns relevant to a new case.Rare or representative cases may retain detailed traces, while common cases may be compressed into pattern-level summaries.
- B Expert Memory Design: Memory fusion grounds stage-specific expert outputs in the current input, stable domain priors, evolving case context, and relevant historical prototypes.The formulation explicitly treats memory as an influence on the final expert-agent representation.
C Energy Setting
The framework assigns clinician-informed fixed energy weights to expert agents to simulate their relative diagnostic burden during routing.
- Energy weights are clinician-informed simulation references for relative specialist burden, not direct estimates of real-world financial or operational cost.The scheme uses triage and fusion agents at 1.0, while specialized agents receive higher weights based on diagnostic burden.
D ClinFEScore
ClinFEScore evaluates clinical conclusions by extracting atomic facts, embedding them in a shared biomedical semantic space, and comparing predicted and reference facts. It reports precision, recall, and overall fact similarity rather than relying on superficial text similarity.
- ClinFEScore converts generated and reference conclusions into sets of atomic clinical facts before semantic comparison.Facts may represent symptoms, imaging findings, pathological clues, or diagnostic conclusions.
- A biomedical text encoder maps each extracted fact into a shared semantic vector space for comparison.Pairwise fact proximity is measured with cosine similarity.
- Fact Precision measures how well facts expressed in the generated conclusion are supported by the reference answer.It averages each predicted fact’s best semantic match against the reference fact set.
- Fact Recall measures how well clinically relevant facts in the reference answer are covered by the generated conclusion.It averages each reference fact’s best semantic match against the predicted fact set.
- Fact Similarity measures overall semantic similarity between the generated and reference fact sets after concatenating each set into a global summary.The summaries are encoded by the same biomedical text encoder.
- ClinFEScore distinguishes factual support from factual coverage instead of measuring only superficial similarity between free-text conclusions.Its design combines fact-aware semantic comparison with separate precision and recall perspectives.
E Routing Layers Ablation
The routing-layer ablation results are reported in Table 3.
- Table 3 provides the detailed routing-layer ablation results.