Source-linked AI summary
GraM-Diff: A Unified Graph-Mamba Diffusion Framework for EEG-Based Alzheimer's Disease Data Generation and Diagnosis
M. Tanveer, Ayush Singh Rana, Sanskriti Jain, Arnav Kumar, Aryaman Tiwari, A. Rahaman, A. Quadir, M. Sajid
TL;DR
Small, imbalanced EEG datasets and fragmented or structurally limited generative methods hinder AD data generation and diagnosis. GraM-Diff addresses this with a unified classifier-guided Graph-Mamba diffusion model that combines graph spatial encoding, linear-time temporal modeling, and latent guidance. Across four EEG-based AD benchmarks, synthetic augmentation improves classification and distributional quality while retaining robustness under data scarcity and noise.
Problem
Small, imbalanced AD-EEG datasets and existing class-specific or structurally limited generative approaches constrain effective EEG synthesis and diagnosis.
Method
GraM-Diff uses a unified classifier-guided diffusion model with GCN spatial encoding, Bidirectional Mamba temporal modeling, and latent-space conditioning for shared healthy and pathological EEG generation.
Results
Synthetic augmentation improves downstream AD classification and produces superior Context-FID and correlation scores over strong generative baselines across four EEG-based AD benchmarks.
Takeaways & Limitations
A single shared-latent generator can produce diagnostically useful multichannel EEG and support classification in data-scarce settings with robustness under moderate noise.
Takeaways & Limitations
The reported experiments use a fixed fully connected adjacency matrix rather than dataset-specific electrode geometries.
Abstract
from arXiv · showhide
Electroencephalography (EEG) is a promising, non-invasive, and cost-effective modality for Alzheimer's disease (AD) detection, but deep learning methods are limited by small and imbalanced clinical datasets. Generative augmentation offers a solution, yet existing approaches rely on inefficient class-specific models or fail to capture complex spatial and temporal brain dynamics. To address this, we propose GraM-Diff, a unified classifier-guided Graph-Mamba diffusion framework for EEG synthesis. It embeds Graph Convolutional Networks within a diffusion U-Net to model inter-electrode connectivity and Bidirectional Mamba state-space blocks for linear-complexity long-range temporal modeling. Latent-space classifier guidance lets a single model generate both healthy and pathological EEG within a shared representation, avoiding fragmented per-cohort pipelines. Across four EEG-based AD benchmarks, synthetic augmentation improves classification, yields superior Context-FID and correlation scores over strong generative baselines, and enhances robustness in data-scarce settings.
1 Introduction
EEG offers accessible, high-temporal-resolution support for AD assessment, but scarce, imbalanced datasets and fragmented generative approaches constrain deep-learning development. GraM-Diff addresses these challenges with one classifier-guided generator that models spatial and temporal EEG structure.
- EEG is non-invasive, cost-effective, and suited to longitudinal assessment because of its accessibility and high temporal resolution.
- Annotated EEG remains scarce, while small public corpora, class imbalance, and privacy-driven data silos hinder generalization.
- Existing generative approaches often train separate healthy and pathological models, increasing overhead and reducing flexibility.
- GraM-Diff combines topology-aware spatial modeling and long-range temporal modeling in a single generative architecture.Its classifier-guided DDPM conditions generation on diagnostic state within a shared representation rather than using per-cohort generators.
- The framework integrates graph-based spatial priors, Bidirectional Mamba blocks with O(L) temporal cost, and noise-robust LogRoPE guidance.
2 Related Work
Prior EEG-AD models often overlook brain-connectivity topology or incur high sequence-modeling costs, while existing generative methods remain limited by data scarcity and weak disease conditioning. GraM-Diff combines graph modeling, linear-time temporal processing, and classifier guidance to address these gaps.
- Many discriminative EEG-AD models treat channels as independent sequences or grid images, ignoring explicit graph topology.
- Transformer-based EEG-AD models capture multiple time scales but incur quadratic complexity on long recordings and have not been used generatively.
- Public AD-EEG corpora often contain fewer than 330 subjects, making high-capacity training prone to overfitting.
- GAN augmentation can improve accuracy but suffers training instability and mode collapse, while usually synthesizing features instead of raw multichannel signals.
- Existing EEG diffusion models generally lack AD-pathology conditioning and classifier guidance for retaining disease-specific signatures.
- Mamba provides linear-time sequence modeling, and GraM-Diff uses a bidirectional block to capture causal and anti-causal dependencies in extended EEG recordings.
3 Proposed Methodology
GraM-Diff processes multichannel EEG through graph-based spatial encoding, a U-Net temporal encoder–decoder, a bidirectional Mamba bottleneck, and latent-space classifier guidance. The design preserves spatial, temporal, trend, seasonal, and spectral structure during conditional diffusion synthesis.
- Overall architecture: GraM-Diff combines spatial electrode relationships, long-range temporal modeling, and latent-space guidance to generate Alzheimer’s stage-conditioned EEG.
- Diffusion process: The diffusion process corrupts clean EEG X0 through a Gaussian Markov chain, while the denoising network operates on U-Net latent features ht.The process uses T=1000 timesteps and predicts the clean signal before analytically recovering noise for guided sampling.
- Diffusion process: A cosine noise schedule is adopted because linear schedules abruptly destroy fine-grained high-frequency EEG structure.
- Topology-aware spatial encoding: The GCN treats electrodes as nodes and supplies spatially enriched features to the U-Net, while experiments use a fixed fully connected adjacency across datasets.
- Hierarchical temporal encoder–decoder: The U-Net uses four downsampling levels, skip connections, timestep conditioning, and mirrored decoding to reconstruct the denoised signal as trend plus seasonal components.
- Mamba bottleneck: A Bidirectional Mamba bottleneck replaces self-attention, modeling long-range dependencies with O(L) rather than O(L^2) complexity.Its adaptive step size and bidirectional scans support both fast transients and slow background rhythms.
- Latent-space guidance: Latent-space guidance uses overlapping multigranular patches, LogRoPE, and a time-conditioned Transformer classifier to guide diffusion toward diagnostic patterns.
- Training objective: A dual-domain loss combines time-domain reconstruction with frequency-domain spectral constraints to preserve oscillatory structure.The spectral term uses λ=0.1, and direct X0 prediction supports training stability and spectral preservation.
4 Experiments and Results
Experiments evaluate GraM-Diff on three public EEG datasets using subject-independent synthetic-data training, generative baselines, fidelity metrics, utility tests, ablations, and noise sensitivity. The results show strong diagnostic performance, closer agreement with real EEG dynamics, preserved utility under synthetic mixing, component contributions, and graceful degradation under perturbation.
- Experimental setup: Evaluation uses three public EEG datasets with subject-independent 80–20 splits, ∼1.5× synthetic samples, and identical held-out real test data.The diffusion model and downstream classifiers were trained separately under the reported experimental protocol.
- Diagnostic reliability: 97.14% ADSZ accuracy from synthetic-only ADFormer training exceeded its 93.08% real-data result and competing architectures.The same synthetic-training evaluation reached 51.22% on the harder ADFD 3-Class task.
- Generative comparison: 97.18% ADSZ accuracy exceeded Diffusion-TS at 87.41%, TimeGAN at 69.81%, and TimeVAE at 64.25% under identical downstream conditions.The comparison held the classifier constant and evaluated synthesis quality across Accuracy, Precision, F1, AUROC, and AUPRC.
- Generative fidelity: Context-FID and Correlation Scores were lower than Diffusion-TS on every dataset, indicating closer agreement with real EEG dynamics.For ADSZ, Context-FID was 0.109 vs. 0.585 and Correlation Score was 0.275 vs. 0.492.
- Utility and augmentation: 96.87% ADSZ accuracy remained at 80% synthetic data, while imputation MSE reached 0.108 at 90% masking and forecasting MSE was 0.0501 at horizon 32.These tests covered imputation, forecasting, and real/synthetic mixture classification.
- Ablation and sensitivity: Removing Patch Embeddings + LogRoPE reduced ADSZ accuracy from 97.18% to 91.81% and APAVA accuracy from 82.81% to 69.65%.The GNN and Mamba modules each contributed additional gains, while 10% noise left ADSZ AUROC at 85.35% versus 97.82% baseline.
5 Conclusion
GraM-Diff is presented as a unified classifier-guided Graph–Mamba diffusion framework for conditional multichannel EEG synthesis. Its shared latent-space generator combines topology-aware spatial priors with linear-complexity long-range temporal modeling, producing samples that better match real distributions and improve downstream AD classification.
- Conclusion: GraM-Diff couples topology-aware spatial priors with linear-complexity long-range temporal modeling for conditional multichannel EEG synthesis.The framework uses a single classifier-guided model for healthy and pathological signals in a shared latent space.
- Conclusion: The framework produces high-fidelity samples, improves downstream AD classification, and shows graceful robustness under noise.The conclusion also reports superior Context-FID and correlation scores relative to the evaluated generative alternatives.
- Future work: Future work will extend the framework to multimodal and irregularly sampled time series while improving scalability and controllability for clinical deployment.These extensions are stated as future directions rather than evaluated results.