LLM as Clinical Graph Structure Refiner: Improving EEG-Based Seizure Detection

Forum topic · 小凯 · 2026-05-02

Summary

A paper posted on zhichai.net (arXiv: 2604.28178) by Zihan Liao, Mohamed Osama Ahmed, Jinzhuo Luo et al. proposes using large language models (LLMs) to refine graph structures for EEG-based automated seizure detection. EEG signals are noisy, and existing graph construction methods—whether correlation-based or learning-based—often produce redundant or irrelevant edges that degrade representation quality and downstream performance. The authors present a two-stage framework: first they validate that LLM-based edge refinement can identify and remove redundant connections, improving seizure detection accuracy. They then build a robust pipeline where an initial graph is constructed using a Transformer-based edge predictor and an MLP that assign probability scores to potential edges, after which the LLM acts as an edge-set refiner, validating remaining connections using textual and statistical features of node pairs. Extensive experiments on the TUSZ dataset show the framework improves task performance and yields cleaner, more interpretable graph representations, demonstrating the potential of LLMs as reasoning-based structure refiners in clinical graph learning.

Paper Overview

Field: AI / Medical
Authors: Zihan Liao, Mohamed Osama Ahmed, Jinzhuo Luo et al.
Published: 2026-04-30
arXiv: 2604.28178

Motivation

Electroencephalogram (EEG) signals are vital for automated seizure detection, but their inherent noise makes robust representation learning challenging. Existing graph construction methods, whether correlation-based or learning-based, often generate redundant or irrelevant edges due to the noisy nature of EEG data. This significantly impairs the quality of graph representations and limits downstream task performance.

Approach

Motivated by the remarkable reasoning and contextual understanding capabilities of large language models (LLMs), the authors explore using an LLM as a graph edge refiner in a two-stage framework:

1. Feasibility validation: They first show that LLM-based edge refinement can effectively identify and remove redundant connections, significantly improving seizure detection accuracy and producing more meaningful graph structures. 2. Robust pipeline: An initial graph is built with a Transformer-based edge predictor and a multi-layer perceptron (MLP), which assign probability scores to potential edges; a threshold determines edge existence. The LLM then serves as an edge-set refiner, making informed decisions to validate remaining connections based on the textual and statistical features of node pairs.

Results

Extensive experiments on the TUSZ dataset demonstrate that the LLM-refined graph learning framework:

Improves seizure detection task performance.
Produces cleaner, more interpretable graph representations.

Source

arXiv:2604.28178

*Auto-collected on 2026-05-02.*