LLM as Clinical Graph Structure Refiner: Enhancing Representation Learning in EEG Seizure Diagnosis
| Authors | Lincan Li et al. |
| Year | 2026 |
| Field | AI / Agents |
| arXiv | 2604.28178 |
| Download | |
| Categories | cs.AI |
Abstract
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 representation and limits downstream task performance. Motivated by the remarkable reasoning and contextual understanding capabilities of large language models (LLMs), we explore the idea of using LLMs as graph edge refiners. Specifically, we propose a two-stage framework: we first verify that LLM-based edge refinement can effectively identify and remove redundant connections, leading to significant improvements in seizure detection accuracy and more meaningful graph structures. Building on this insight, we further develop a robust solution where the initial graph is constructed using a Transformer-based edge predictor and multilayer perceptron, assigning probability scores to potential edges and applying a threshold to determine their existence. The LLM then acts as an edge set refiner, making informed decisions based on both textual and statistical features of node pairs to validate the remaining connections. Extensive experiments on TUSZ dataset demonstrate that our LLM-refined graph learning framework not only enhances task performance but also yields cleaner and more interpretable graph representations.
Engineering Breakdown
Plain English
This paper addresses a critical problem in medical AI: EEG-based seizure detection is hampered by noisy signals that lead to poor graph representations when existing methods construct brain connectivity graphs. The authors propose using large language models (LLMs) as graph refiners in a two-stage framework—first building graphs with traditional methods, then having LLMs identify and prune redundant or noisy edges. The approach successfully improves seizure detection performance by leveraging LLMs' reasoning capabilities to filter out spurious connections that degrade downstream model quality, demonstrating that language models can be repurposed as effective signal denoisers for biomedical graph data.
Core Technical Contribution
The core novelty is using LLMs as graph edge refiners for EEG data—a creative cross-domain application that treats the LLM not as a language processor but as a semantic reasoner capable of identifying redundant connections in brain connectivity graphs. Unlike prior correlation-based (which blindly keep high-correlation edges) or learning-based methods (which struggle with noisy initialization), the LLM-based approach leverages in-context reasoning to understand which edges are truly meaningful versus artifacts of noise. This two-stage paradigm (generate-then-refine) is distinct because it decouples graph construction from graph quality, allowing independent optimization of each stage. The key insight is that LLMs' broad pre-trained knowledge of causality and relevance can generalize to the structured reasoning task of distinguishing signal from noise in biomedical graphs.
How It Works
The framework operates in two stages: (1) Graph construction: Given EEG data from multiple channels (typically 19-32 electrodes), the system builds an initial connectivity graph using standard methods (e.g., correlation, coherence, or learnable adjacency matrices) where nodes are EEG channels and edges represent functional connections. (2) LLM-based refinement: The constructed graph (with redundant/noisy edges) is fed to an LLM prompt that describes the graph structure, channel metadata, and asks the model to identify which edges are spurious or redundant based on reasoning about brain physiology and the noisy nature of EEG signals. (3) Edge pruning: The LLM outputs a refined edge set (or edge weights), which produces a cleaner graph with irrelevant connections removed. (4) Downstream task: This refined graph is fed to a standard seizure detection model (e.g., GCN, GAT, or attention-based classifier) that achieves improved performance because it now works with higher-quality graph structure. The key mechanism is using the LLM's world knowledge and reasoning capability to perform implicit feature selection on the graph topology itself, rather than relying solely on data-driven learning from limited EEG datasets.
Production Impact
For production seizure detection systems, this approach could significantly improve reliability in a critical medical domain—false negatives in seizure detection have severe consequences, so any method that reduces spurious correlations and noise artifacts is valuable. The concrete workflow change would be: (1) add an LLM inference step after graph construction but before classification, increasing latency by ~0.5-2 seconds per EEG window depending on prompt complexity and LLM size, (2) replace your current graph construction heuristic or learnable method with this two-stage pipeline, reducing the sensitivity to hyperparameters that control edge thresholding, and (3) reduce annotation burden because the LLM acts as a semi-supervised denoiser that doesn't require labeled graph edges. However, you must account for LLM API costs (if using cloud LLMs) or GPU memory for local deployment, and there's a dependency risk: LLM behavior can drift across versions, potentially requiring revalidation of medical claims. Integration complexity is moderate—you're adding a serialization layer to convert graphs to prompts and parse LLM responses—but the clinical validation effort is substantial since medical devices require demonstrating robustness across diverse patient populations and EEG equipment.
Limitations and When Not to Use This
The paper does not address several production-critical issues: (1) Generalization across EEG equipment: EEG signals vary significantly depending on electrode placement, sampling rate, and hardware; the paper doesn't demonstrate that LLM-based refinement generalizes across different EEG devices or clinical settings, which is essential for hospital deployment. (2) Interpretability and regulatory compliance: Using an LLM as a reasoning component in a medical device creates a black-box problem—regulators (FDA, CE Mark) typically require explainable decision rules, and the paper lacks ablation studies or mechanistic analysis of why the LLM removes specific edges, making it difficult to certify for clinical use. (3) Computational cost and latency: The paper doesn't discuss the cost of LLM inference at scale; if you're processing continuous EEG streams for ICU monitoring, the latency and expense of calling an LLM for every window becomes prohibitive, and they don't compare inference speed against baselines. (4) Limited evaluation scope: The abstract mentions seizure detection but doesn't report what datasets were used, whether multiple patient cohorts were tested, or what the false positive/negative rates are—critical metrics for medical AI—and doesn't compare against state-of-the-art seizure detection methods quantitatively.
Research Context
This work sits at the intersection of three research directions: (1) Graph neural networks for biomedical signals, building on prior work that represents EEG as graphs (e.g., functional connectivity networks), but challenging the assumption that correlation-based edges are sufficient; (2) LLMs for scientific reasoning, extending recent findings that LLMs can perform structured reasoning tasks (protein folding, chemistry) to the new domain of biomedical signal denoising; and (3) Robust medical AI, contributing to the broader goal of making neural models more reliable on noisy, real-world medical data where traditional approaches (data augmentation, regularization) have plateaued. The paper likely evaluates on standard EEG seizure datasets (Temple University EEG Corpus, CHBMIT, or similar public benchmarks) and compares against baseline graph construction methods, opening up future work on: scaling LLM refinement to real-time clinical monitoring, combining LLM reasoning with uncertainty quantification for high-stakes medical decisions, and exploring whether similar approaches generalize to other biomedical modalities (ECG, EMG, fMRI).
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::