A Clinical Guideline-Grounded Hybrid Agentic Framework for Holistic Epilepsy Management.

This paper proposes a clinical guideline-grounded hybrid multi-agent framework that integrates modality-specific discriminative and generative models with a central orchestrating agent to enable holistic, evidence-informed epilepsy management across six key clinical tasks, demonstrating superior performance over conventional baselines on a newly introduced multi-modal benchmark.

The Gist

Imagine you are trying to solve a very complex mystery: Epilepsy.

Epilepsy isn't just one thing; it's a chaotic puzzle involving electrical storms in the brain (EEG), structural maps of the brain (MRI), genetic clues, and a patient's personal history. In the real world, solving this puzzle requires a team of experts: a neurologist, a radiologist, a surgeon, and a psychiatrist, all sitting around a table, arguing, comparing notes, and consulting a massive rulebook (medical guidelines) to decide the best treatment.

For a long time, Artificial Intelligence (AI) tried to solve this, but it was like having a team of siloed specialists who never talk to each other.

• One AI was great at reading EEGs but knew nothing about MRIs.
• Another AI was great at reading MRIs but couldn't understand the patient's history.
• A third AI (a "Generative" one) could write a nice story about the patient but often made things up (hallucinations) or got confused by the numbers.

The authors of this paper, EPI-GUIDE, decided to build a super-team that actually works together, just like a real hospital team.

The "Super-Team" Analogy

Think of the EPI-GUIDE framework as a high-tech hospital command center with three distinct roles:

1. The "Fact-Checkers" (Discriminative Models)

These are the AI specialists who are incredibly good at spotting specific patterns in data, but they are a bit robotic.

• The Analogy: Imagine a team of forensic accountants. They look at the numbers (EEG waves, MRI pixels) and say, "There is a 90% chance this is a tumor," or "This seizure started in the left temple." They are precise, reliable, and don't make things up, but they can't write a report or explain why in a human way.
• In the paper: These are the models trained specifically on EEGs, MRIs, and clinical text to give hard numbers and probabilities.

2. The "Storytellers" (Generative Agents)

These are the AI specialists who can read the data and write a narrative, but they can sometimes get the facts wrong.

• The Analogy: Imagine a team of creative journalists. They look at the same data and say, "It looks like the patient is having a focal seizure, and here is a story about why." They are great at connecting dots and explaining things, but without the "accountants," they might invent a detail that isn't there.
• In the paper: These are Large Language Models (LLMs) that generate textual interpretations of the medical images and records.

3. The "Chief Detective" (The Orchestrating Agent)

This is the most important part. This AI is the Team Leader who sits in the middle.

• The Analogy: Imagine a seasoned Police Chief who has the Official Rulebook (International Epilepsy Guidelines) open on their desk.
  • The Chief listens to the "Accountants" (Fact-Checkers) for the hard numbers.
  • The Chief listens to the "Journalists" (Storytellers) for the narrative.
  • The Magic: If the Accountant says "Left Side" and the Journalist says "Right Side," the Chief doesn't just guess. They open the Rulebook, check the official medical guidelines, and say, "According to the rules, if the EEG shows X and the MRI shows Y, we must conclude Z."
  • If the evidence is messy, the Chief asks for more info (a "Follow-up") before making a final decision.

Why is this a big deal?

The Problem with the Old Way:

• Siloed AI: Like having a doctor who only looks at your feet and ignores your heart.
• Pure "Chatbot" AI: Like having a doctor who is very confident but makes up facts because they are trying to be helpful.

The EPI-GUIDE Solution:
By combining the hard facts (Discriminative) with the narrative reasoning (Generative) and forcing them to follow the official rulebook (Guidelines), the system becomes:

1. More Accurate: It got about 86% accuracy on complex tasks, beating the previous best methods by a significant margin.
2. More Reliable: It doesn't just guess; it checks its work against medical rules.
3. Holistic: It looks at the whole patient, not just one scan.

The Result

The researchers tested this "Super-Team" on two different groups of patients. In every test, the team that used the Chief Detective with the Rulebook (EPI-GUIDE) solved the mystery better than any single specialist or any team that didn't have the rulebook.

In short: They built an AI that doesn't just "know" things; it thinks like a real medical team, checks its work against the rulebook, and makes decisions that are safe, accurate, and ready for the real world.

Technical Summary

1. Problem Statement

Epilepsy management is inherently multidisciplinary, requiring the integration of heterogeneous data (EEG, MRI, genomics, clinical history) to perform complex tasks such as seizure detection, syndrome diagnosis, epileptogenic zone (EZ) localization, and surgical outcome prediction.

• Current Limitations: Existing AI approaches are typically "siloed," meaning they are trained for single tasks on specific modalities (e.g., an EEG model for seizure detection only). They fail to replicate the collaborative workflow of human specialists who synthesize diverse evidence.
• Generative AI Limitations: While Large Language Models (LLMs) and foundation models offer reasoning capabilities, purely generative systems often suffer from hallucinations, instability in numerical reasoning, and a lack of grounding in quantitative, modality-specific signals.
• The Gap: There is a need for a unified framework that combines the reliability of discriminative models with the reasoning capabilities of generative agents, all grounded in established clinical guidelines.

2. Methodology: EPI-GUIDE

The authors propose EPI-GUIDE, a hybrid multi-agent framework designed to emulate the workflow of a multidisciplinary epilepsy team. The system consists of three core components:

A. Modality-Specific Evidence Generation

The framework processes patient data through two parallel pathways for each modality (e.g., EEG, MRI, Clinical Text):

1. Discriminative Models: Specialized supervised models (e.g., ResNet-50, MedSigLIP, PubMedBERT, XGBoost) generate structured predictions (class probabilities, risk scores). These outputs are converted into textual evidence ($\tilde{s}^{(d)}_i$) to serve as auxiliary guidance.
2. Generative Models: Modality-specific LLM agents (e.g., MedGemma) analyze the raw inputs to produce textual clinical interpretations and structured predictions ($\tilde{s}^{(g)}_i$) with justifications.

• Aggregation: The structured predictions and textual interpretations are combined into a unified multimodal evidence set ($E$).

B. Guideline-Grounded Orchestrating Agent

A central "Orchestrator" agent manages the decision-making process:

• Retrieval-Augmented Generation (RAG): The agent queries a curated knowledge base of 66 international epilepsy guidelines (ILAE, NICE, textbooks) to retrieve relevant clinical pathways ($G$) based on the specific task.
• Reasoning & Coordination: The agent receives the aggregated evidence ($E$) and the retrieved guidelines ($G$). It performs multi-turn interactions to:
  • Identify concordance or discordance between modalities.
  • Resolve conflicts using the retrieved guidelines.
  • Decide whether to finalize the prediction (COMPLETE) or request further targeted analysis (FOLLOW-UP).
• Output: The system produces final clinical recommendations (e.g., epilepsy type, surgical candidacy) grounded in evidence and guidelines.

C. Implementation Details

• Models: Uses MedGemma (27B and 4B) for generation, and a mix of CNNs (ResNet, MedSigLIP) and Transformers (PubMedBERT, REVE-base) for discrimination.
• Orchestrator: Powered by GPT-OSS-120b with a temperature of 0.1 for deterministic reasoning.
• Hardware: Trained and evaluated on 8×A100-80GB GPUs.

3. Key Contributions

1. First Agentic Framework for Holistic Epilepsy Care: Introduces EPI-GUIDE, the first system to apply a hybrid multi-agent approach to comprehensive epilepsy management across multiple tasks.
2. Hybrid Paradigm: Successfully couples the robustness of discriminative models (for quantitative accuracy) with the reasoning and flexibility of generative LLMs.
3. Clinical Grounding: Integrates international clinical guidelines directly into the agent's reasoning loop via RAG, ensuring decisions are evidence-informed and clinically compliant.
4. New Benchmark: Introduces a publicly available, curated Multi-modal Multi-task Epilepsy (MME) dataset containing 306 patients with annotations for five distinct management tasks, alongside benchmarks on a public HD-EEG dataset.

4. Results

The framework was evaluated on two datasets: the curated MME dataset (5 tasks) and the public HD-EEG dataset (2 tasks).

• Performance on MME Dataset:
  • EPI-GUIDE achieved an overall accuracy of 85.8%.
  • It outperformed the best discriminative ensemble (78.7%) by +7.1% and the best generative baseline (MedGemma-27B, 61.8%) by +24.0%.
  • Significant improvements were observed in data-scarce tasks like Surgical Outcome prediction (+5.7% over ensemble).
• Performance on HD-EEG Dataset:
  • EPI-GUIDE achieved 81.3% overall accuracy, surpassing the supervised foundation model (REVE-base, 74.5%) and a Multi-Agent LLM baseline (73.0%).
  • The most significant gain was in EZ localization (+14.9% over the Multi-Agent LLM).
• Ablation Studies:
  • Removing discriminative guidance caused a massive drop in accuracy (85.8% $\to$ 62.9%), proving the necessity of structured evidence.
  • Removing guideline grounding (RAG) caused a 7.0% drop, highlighting the importance of clinical protocols.
• Expert Evaluation:
  • In a qualitative study with neurologists, EPI-GUIDE achieved 93.4% task-level accuracy compared to 54.1% for an LLM-only baseline.
  • It received a reliability rating of 3.64/5 vs. 2.64/5 for the LLM-only approach.

5. Significance

This work represents a paradigm shift in medical AI from isolated, single-task models to holistic, guideline-grounded agentic systems.

• Clinical Relevance: By mimicking the multidisciplinary team workflow and adhering to international guidelines, the system offers a more trustworthy and clinically viable tool for decision support.
• Robustness: The hybrid approach mitigates the hallucination risks of pure LLMs while overcoming the "black box" limitations of pure discriminative models.
• Future Impact: The release of the MME dataset and code provides a critical resource for advancing research in multi-modal, multi-task medical AI, potentially applicable to other complex chronic diseases requiring integrated care.