Empowering Locally Deployable Medical Agent via State Enhanced Logical Skills for FHIR-based Clinical Tasks

This paper introduces SELSM, a training-free framework that enhances locally deployable medical agents by distilling simulated clinical trajectories into entity-agnostic logical rules, thereby significantly improving zero-shot FHIR-based task performance and achieving a 100% completion rate on the MedAgentBench without compromising data privacy.

The Gist

The Big Problem: The "Super-Doctor" Who Can't Read the Map

Imagine you have a brilliant, super-smart AI doctor (a Large Language Model). It knows everything about medicine, can diagnose diseases, and understand complex medical jargon. You want to put this AI in a hospital to help real doctors.

But there's a huge catch: Hospitals are like different video games with different control schemes.

• Hospital A uses a specific digital system to order tests.
• Hospital B uses a totally different system to check patient allergies.
• Hospital C has its own unique way of scheduling appointments.

To teach the AI how to use these specific systems, you usually need to show it thousands of examples of real doctors using them. But here's the problem: Real patient data is top-secret. You can't just copy-paste it to a cloud server to train the AI because that violates privacy laws. Plus, hospitals can't afford to buy super-expensive computers to retrain the AI for every single hospital.

So, the AI is stuck. It's smart, but it doesn't know how to press the right buttons in this specific hospital. If you ask it to order an MRI, it might guess the wrong format, crash the system, or get lost.

The Solution: SELSM (The "Universal Rulebook")

The authors of this paper created a new framework called SELSM (State-Enhanced Logical-Skill Memory). Think of it not as teaching the AI new facts, but giving it a Universal Rulebook and a Smart Librarian.

Here is how it works, broken down into three simple steps:

1. The "Simulator" (Learning without Risk)

Instead of using real patients, the team built a video game simulation of a hospital. They let the AI play the game, trying to order meds, check records, and schedule tests.

• The Magic Trick: When the AI makes a mistake or succeeds in the game, a "Judge" (another AI) doesn't just say "Good job" or "Bad job." Instead, it strips away the specific details (like "Patient John Doe" or "Hospital 5") and writes down the general logic.
• The Analogy: Imagine learning to drive. Instead of memorizing "Turn left at the red barn on Main Street," you learn the rule: "When the light is green and the road is clear, turn left." The AI learns these "rules of the road" without ever needing to see a real car or a real street.

2. The "Universal Rulebook" (Entity-Agnostic Skills)

The AI collects these general rules into a library. These rules are Entity-Agnostic, which is a fancy way of saying they don't care about who the patient is or which hospital it is.

• The Analogy: It's like a recipe book. A recipe for "Scrambled Eggs" works whether you are cooking in a kitchen in New York or a kitchen in Tokyo. You don't need a new recipe for every single egg; you just need the logic of how to scramble them.

3. The "Smart Librarian" (Two-Stage Retrieval)

When the AI is actually working in a real hospital (or a new simulation), it doesn't guess what to do. It asks its Smart Librarian.

• Step 1 (The Topic): The AI says, "I need to order an MRI." The Librarian looks at the "Topic" and pulls out the "Medical Imaging" section of the Rulebook.
• Step 2 (The Context): The AI says, "Okay, but the system just told me the patient has an allergy." The Librarian looks at the current situation and finds the specific rule: "If allergy exists, check the alternative drug list first."
• The Result: The AI gets a hint on exactly what to do next, step-by-step, without ever changing its own brain (its code).

Why This is a Game-Changer

The paper tested this on a very tough benchmark (MedAgentBench) that mimics real hospital chaos. Here is what happened:

• Before (The Baseline): The AI was smart but clumsy. It often forgot steps, got confused by the hospital's specific formatting, or gave up entirely. It succeeded only about 41% of the time.
• After (With SELSM): The AI became a master navigator. It succeeded 61% of the time overall.
• The Star Performer: On one specific model (Qwen3-30B), the AI achieved a 100% completion rate. It never got stuck, never crashed, and finished every single task.

The "Secret Sauce" Analogy

Imagine you are trying to solve a maze.

• Old Way: You try to memorize the entire maze by heart. If the maze changes even a little (a wall moves), you get lost.
• SELSM Way: You are given a compass and a map of the rules (e.g., "Always turn right at dead ends," "Don't go into water"). You don't memorize the maze; you memorize the logic of navigating mazes. No matter what the maze looks like, you can find your way out.

The Bottom Line

This paper proves that we don't need to break privacy laws or spend millions on supercomputers to make AI doctors useful. Instead, we can teach them the logic of how hospitals work using safe simulations, and then give them a "cheat sheet" of those rules when they go to work.

It turns a "smart but clueless" AI into a "smart and adaptable" agent that can walk into any hospital, read the local rules, and get the job done safely.

Technical Summary

1. Problem Statement

The deployment of Large Language Model (LLM) agents in clinical settings faces a critical bottleneck: data scarcity under strict privacy constraints.

• Privacy Barriers: Clinical interaction trajectories are highly sensitive and cannot be transmitted to cloud-based models for training.
• Resource Constraints: Acquiring annotated multi-turn data for Reinforcement Learning (RL) is prohibitively difficult, and ultra-large models are computationally too expensive for most local hospital deployments.
• Brittleness of Current Approaches: Traditional methods that embed environment-specific workflows into model weights (via fine-tuning or RL) fail in real-world scenarios due to the heterogeneity of hospital IT infrastructures (e.g., different FHIR endpoints, formularies, and protocols). This leads to "state polysemy," where identical intermediate states (e.g., a successful API response) appear in different workflows, causing agents to hallucinate or break task chains when encountering unseen environments.
• Performance Gap: Even state-of-the-art models struggle with zero-shot execution on high-fidelity EHR benchmarks, often failing due to strict formatting constraints and an inability to recover from intermediate API errors.

2. Methodology: SELSM Framework

The authors propose State-Enhanced Logical-Skill Memory (SELSM), a training-free, memory-augmented framework designed to equip locally deployable models (30B–32B parameters) with reusable operational logic without modifying model weights.

The framework operates in three phases:

Phase 1: Logical Skill Distillation

• Trajectory Collection: An agent interacts with a simulated FHIR server (virtual EHR) to generate interaction trajectories ($\tau$).
• LLM-as-Judge Evaluation: A judge model evaluates every step of the trajectory, not just the final outcome. It produces:
  • A binary correctness label ($\ell_t$).
  • A Logical Skill ($e_t$): An entity-agnostic natural language string residing in an abstract skill space ($\mathcal{Z}$).
• Abstraction Protocol: The judge strips specific entities (patient IDs, specific URLs, hard-coded values) to create generalized rules (e.g., "verify identity $\to$ check allergies $\to$ query formulary"). This ensures the skills are transferable across different institutions.
• Dual Strategy: The system generates "Success Paradigms" for correct steps and "Correction Axioms" for errors, encoding both what to do and what to avoid.

Phase 2: Hierarchical Memory Indexing

The distilled skills are organized into a two-level embedding index:

1. Task-Level Index: Embeds the clinical query ($q$) to anchor the retrieval in the correct semantic neighborhood.
2. Transition-Level Index: Embeds the specific state ($s_t$) and the corresponding logical skill ($e_t$) to enable fine-grained matching during execution.

Phase 3: Query-Anchored Two-Stage Retrieval

During inference, when the agent faces a new task, SELSM performs a hierarchical retrieval to resolve state polysemy:

1. Stage 1 (Task-Level Filtering): Scores stored records based on the similarity between the current task query and the task-level index. This filters out irrelevant workflows.
2. Stage 2 (Transition-Level Ranking): From the filtered candidates, it ranks specific transitions based on the similarity between the current full state and the stored state/skill pairs.
3. Skill Injection: The top-ranked logical skills are injected into the agent's prompt as structured prior knowledge, guiding step-by-step reasoning without altering model parameters.

3. Key Contributions

1. Training-Free Framework: SELSM circumvents the need for parameter updates or sensitive data transmission, making it viable for privacy-constrained, locally deployed hospitals.
2. Entity-Agnostic Abstraction: By distilling raw trajectories into abstract logical skills, the framework solves the "state polysemy" problem, allowing models to generalize across heterogeneous clinical environments.
3. Query-Anchored Two-Stage Retrieval: A novel mechanism that prevents the retrieval of operationally irrelevant skills by first anchoring the search to the task intent before matching specific states.
4. Empirical Validation: Extensive evaluation on MedAgentBench, the only high-fidelity virtual EHR sandbox grounded in real clinical data, demonstrating superior zero-shot performance.

4. Experimental Results

The framework was evaluated on MedAgentBench using three locally deployable backbones: Qwen3-30B-A3B, Qwen3-32B, and GLM4-32B.

• Performance Gains:
  • Qwen3-30B-A3B: Achieved a 100% Task Completion (TC) rate (eliminating all chain breakdowns) and boosted the overall Success Rate (SR) by an absolute 22.67% (from 48.66% to 71.33%).
  • GLM4-32B: Showed the most dramatic improvement in Action Success Rate (A-SR), jumping from 16.66% to 60.66% (+44.00%).
  • Average Improvement: Across all models, the average SR increased by 19.67%, and A-SR by 24.89%.
• Comparison with SOTA: SELSM significantly outperformed existing memory-augmented baselines (ExpeL and A-Mem). Notably, A-Mem degraded performance (SR dropped to 30.44%), confirming that retrieving raw, noisy trajectories is harmful in medical contexts, whereas abstracted logical skills are effective.
• Efficiency:
  • One-Shot Correct Rate (OSR): SELSM achieved the highest OSR (71.3% on Qwen3-30B-A3B), indicating the model selects the correct action on the first attempt more often.
  • Token Efficiency: SELSM achieved Pareto-dominant performance, delivering higher success rates with lower token consumption per correct task compared to baselines.
• Failure Mode Analysis: SELSM converted "hard failures" (invalid API calls, timeouts) into "soft failures" (incorrect but completed answers), significantly improving robustness.

5. Significance and Future Directions

• Privacy-Preserving AI: SELSM offers a practical pathway for hospitals to deploy capable AI agents without violating patient privacy or requiring massive computational resources for training.
• Cross-Institutional Generalization: The entity-agnostic design suggests that a hospital can generate its own specific operational skills from a local simulator and deploy them immediately, without retraining the base model.
• Limitations: The study relies on a single benchmark (MedAgentBench) and a simulated environment. The "LLM-as-Judge" quality is bounded by the judge model's capabilities, and the framework currently supports only text-based FHIR interactions, not multimodal workflows (e.g., DICOM images).
• Conclusion: The paper demonstrates that equipping locally deployable models with a dynamically updatable, state-enhanced cognitive scaffold is a viable and efficient strategy for adapting AI to complex, heterogeneous clinical information systems.