EEG-MedRAG: Hypergraph RAG for EEG Data

• EEG-MedRAG is a three-layer hypergraph-based retrieval-augmented generation framework that integrates EEG signals, clinical knowledge, and patient histories into a unified relational structure.
• It employs semantic-temporal joint retrieval and causal-chain diagnostic reasoning to enhance EEG interpretation and support data-driven clinical decision-making.
• Benchmark results show significant gains over traditional methods, with improved diagnostic accuracy and validation across multiple neurological disorders and clinical roles.

EEG-MedRAG is a three-layer hypergraph-based retrieval-augmented generation (RAG) framework designed to address the efficient retrieval and semantic interpretation of large-scale, heterogeneous EEG (electroencephalography) data in clinical and neuroscience contexts. By unifying EEG domain knowledge, patient case histories, and extensive EEG repositories into a single traversable n-ary relational hypergraph, EEG-MedRAG enables semantically coherent, temporally informed retrieval and diagnostic reasoning. It further introduces a comprehensive clinical QA benchmark for evaluating generalization and role-specific reasoning across multiple neurological disorders and professional perspectives (Wang et al., 19 Aug 2025).

1. Hierarchical Hypergraph Structure

EEG-MedRAG operationalizes the complete EEG information space as a hypergraph $G = (V, E)$ with three disjoint but interconnected layers:

• Knowledge Hypergraph (KGH):

This layer encodes domain knowledge as entities ($V^K$) and hyperedges ($E^H$) representing n-ary clinical relations. Each clinical document $d \in \mathcal{D}$ is parsed into facts $F_n^d = \{ (e_i^H, V_i^H) \}$, associating a natural-language relation (hyperedge) $e_i^H$ with involved entity tuples $V_i^H$. The incidence matrix $H^K \in \{0,1\}^{|V^K| \times |E^H|}$ defines the bipartite structure (entities to hyperedges), and all nodes are embedded via a textual encoder $f_\text{text}: V \rightarrow \mathbb{R}^d$.

• Patient Cases Hypergraph (CGH):

Patient cases are modeled as nodes ($V^C$) and multi-attribute hyperedges ($V^K$0). Each case $V^K$1 is represented as a tuple $V^K$2 and embedded via $V^K$3. Augmented pseudo-cases are included through soft nearest-neighbor inference, creating $V^K$4. Incidence matrices are defined analogously to KGH.

• EEG Vector Database (EVD):

Raw EEG segments $V^K$5 are transformed via channel-wise Piecewise Aggregate Approximation (PAA) with window size $V^K$6:

$V^K$7

Each EEG segment is thus represented as a fixed-length embedding (nodes $V^K$8), absent higher-order hyperedges.

The global hypergraph may be represented either as a high-rank incidence tensor $V^K$9 for n-ary relations or as layer-wise incidence matrices. This architecture supports efficient semantic, structural, and temporal traversals.

2. Semantic–Temporal Joint Retrieval

Given a new patient query, consisting of a raw EEG recording $E^H$0 and metadata $E^H$1 (age, symptoms, history), EEG-MedRAG performs multi-step context retrieval:

• EEG-Level Retrieval:

$E^H$2 is computed as above. For each stored EEG vector $E^H$3, the Dynamic Time Warping (DTW) distance is calculated:

$E^H$4

The top-$E^H$5 segments minimizing DTW are retrieved:

$E^H$6

• Hyperedge-Level (Knowledge) Retrieval:

Query metadata is encoded as $E^H$7. For each knowledge hyperedge $E^H$8, with embedding $E^H$9, cosine similarity is computed:

$d \in \mathcal{D}$0

The top-$d \in \mathcal{D}$1 hyperedges are selected:

$d \in \mathcal{D}$2

• Entity-Level Expansion:

For each retrieved hyperedge, incident entities are collected via $d \in \mathcal{D}$3. Additional hyperedges may be fetched for enhanced context.

• Subgraph Fusion:

The three retrieval sets—EEG similarities ($d \in \mathcal{D}$4), knowledge hyperedges ($d \in \mathcal{D}$5), and entities ($d \in \mathcal{D}$6)—are merged. The context subgraph $d \in \mathcal{D}$7 is computed as:

$d \in \mathcal{D}$8

This union defines a connected, clinically meaningful evidence set by closure within one or two hops in the overall hypergraph.

This retrieval process tightly integrates low-level signal similarity, high-level semantic knowledge, and graph connectivity, supporting comprehensive evidence gathering.

3. Causal-Chain Diagnostic Reasoning

Context subgraph $d \in \mathcal{D}$9 forms the prompt for a diagnostic LLM, e.g., GPT-4o-mini. Generation comprises:

• Prompt Assembly:
  • A “diagnostic instruction” template $F_n^d = \{ (e_i^H, V_i^H) \}$0 (“Based on the following EEG segments, patient history, and medical facts, provide a diagnostic reasoning chain and final recommendation.”)
  • Retrieved EEG PAA vectors or their summaries
  • Textual descriptions of entities and hyperedges in $F_n^d = \{ (e_i^H, V_i^H) \}$1
  • The patient’s question $F_n^d = \{ (e_i^H, V_i^H) \}$2
• Conditional Generation:

The model samples the output:

$F_n^d = \{ (e_i^H, V_i^H) \}$3

where $F_n^d = \{ (e_i^H, V_i^H) \}$4 denotes the instruction-tuned LLM.

• Supervised Fine-Tuning (optional):

The approach is amenable to chain-of-thought or stepwise reasoning supervision via the loss:

$F_n^d = \{ (e_i^H, V_i^H) \}$5

where $F_n^d = \{ (e_i^H, V_i^H) \}$6 are LLM parameters.

The explicit fusion of retrieved semantic, structural, and time-series evidence enables interpretable, traceable causal-chain diagnostic outputs, in contrast to prior RAG methods lacking domain-specific graph structure or temporal alignment.

4. Cross-Disease, Cross-Role EEG Clinical QA Benchmark

EEG-MedRAG introduces the first large-scale, systematically curated EEG clinical QA dataset spanning both disorder diversity and professional roles:

EEG Disorder Domain	Clinical Roles	Example Data Sources
Epilepsy	Doctor, Patient, Nurse,	CHB-MIT, OpenNeuro
Depression	Researcher, Hospital Intern	Various OpenNeuro sets
Parkinson’s Disease
Alzheimer’s Disease
Sleep Deprivation
Psychiatric Disorders
mTBI

• Coverage:
  • 7 domains: Epilepsy, Depression, Parkinson’s Disease, Alzheimer’s Disease, Sleep Deprivation, Psychiatric Disorders (e.g., psychosis), Mild Traumatic Brain Injury
  • 5 roles: Doctor, Patient, Researcher, Hospital Intern, Nurse
• Data Construction:
  • EEG data sourced from CHB-MIT (23 epilepsy patients) and seven OpenNeuro datasets (over 200 subjects).
  • Knowledge and QA pairs are derived from clinical guidelines, textbooks, interpretation protocols; totals ~2,500 QA pairs (7 domains × 5 roles × 50–100 questions each).
  • Each QA pair undergoes hierarchical retrieval-based automatic generation, followed by validation by two board-certified neurologists.
• Annotation and Release:

Annotated with salient EEG events and labeled factual relations, role-conditioned prompts, and human-validated ground truths; provided in standardized JSON format, split into train/dev/test without cross-subject overlap.

This dataset enables systematic benchmarking of disease-agnostic reasoning and role-awareness within the EEG context (Wang et al., 19 Aug 2025).

5. Empirical Performance and Ablation Analysis

EEG-MedRAG demonstrates quantifiable gains over established baselines in diagnostic accuracy and retrieval:

• Metrics:
  • Exact Match (EM): binary correctness
  • Token-level F1: overlap between generated and reference responses
• Baselines Compared:
  • NaiveGeneration (LLM without retrieval)
  • StandardRAG (chunk-based retrieval)
  • HyperGraphRAG (binary-relation graph RAG)
  • TimeRAG (time-series retrieval, no domain knowledge)
• Key Results (GPT-4o-mini backbone):
  • F1:
  • NaiveGeneration: 43.35
  • StandardRAG: 45.44
  • HyperGraphRAG: 48.93
  • TimeRAG: 48.21
  • EEG-MedRAG: 53.16 (+7.72 vs. StandardRAG)
  • EM:
  • StandardRAG: 26.19
  • EEG-MedRAG: 32.16 (+5.97)
  • Domain-specific improvements >10 F1 points (e.g., Alzheimer’s 53.18→68.06)
  • Similar performance improvements observed for Deepseek-r1 (+17.23 F1, +9.49 EM) and Gemini-2.5-flash (+5.35 F1).
• Ablation Study:
  • Full F1: 53.16
  • w/o CL: 49.32 (–3.84)
  • w/o IL: 48.52 (–4.64)
  • w/o EL: 48.26 (–4.90)

These findings indicate that all major components are essential for high performance, and the multi-layer hypergraph structure is critical for semantic, structural, and temporal integration.

6. Implementation and Public Resources

EEG-MedRAG is released as open-source software with all code and data available at https://github.com/yi9206413-boop/EEG-MedRAG under the MIT license (data: CC-BY 4.0).

• Codebase Organization:
  • /preprocess: EEG PAA segmentation and embedding
  • /graph_build: Hypergraph (KGH/CGH) construction
  • /retrieval: DTW-based EEG and hyperedge retrieval
  • /generation: Prompt assembly and LLM interfaces
  • /benchmark: Dataset loaders and evaluation scripts
• Dependencies:

Python 3.8+, PyTorch, Transformers, numpy, scipy, dtw-python, networkx, scikit-learn

• Workflow (Quick Start):

1. Clone repository and install dependencies
2. Download EEG data to /data
3. Run EEG embedding script
4. Build hypergraphs (KGH/CGH)
5. Run retrieval and generation pipeline

This infrastructure provides a standardized and extensible platform for EEG-based, retrieval-augmented clinical reasoning research.

---

EEG-MedRAG’s principled integration of semantic knowledge, temporal dynamics, and clinical context within a hierarchical hypergraph framework advances the field of EEG interpretation and diagnostic QA, with robust empirical validation across multiple disorders and professional perspectives (Wang et al., 19 Aug 2025).