PaperlayerView exact revision on arXiv

Source-linked AI summary

M-LINKX: Multiview Graph Learning for Brain Cognitive Disease Detection

An Phan, Yufei Jin, Xingquan Zhu

arXiv:2608.14847v1cs.LG

TL;DR

EEG-based dementia classification is difficult because signals are noisy, non-stationary, and subject-dependent, while a single connectivity graph may incompletely represent neural interactions. M-LINKX combines multiple functional-connectivity graph views with separate modeling of node features and adjacency information, achieving the best subject-level performance in the main experiments on two three-class EEG datasets.

  • Problem

    EEG-based dementia classification needs accessible tools, but noisy signals and reliance on a single connectivity graph limit representation of neural interactions.

  • Method

    M-LINKX represents fixed-length EEG segments with channel-level node features and multiple functional-connectivity graph views, separately modeling node and adjacency information before weighted fusion.

  • Results

    M-LINKX achieves the best subject-level performance under the main experimental settings on two dementia-related EEG datasets with different diagnostic groups.

  • Takeaways & Limitations

    Multi-view functional-connectivity learning can improve EEG segment-based dementia classification when connectivity information is modeled appropriately.

Abstract

from arXiv · show

Electroencephalogram (EEG) is a non-invasive and relatively low-cost procedure that measures brain electricity for the detection of cognitive diseases. EEG-based classification of dementia-related conditions, including Alzheimer's disease (AD), mild cognitive impairment (MCI), and frontotemporal dementia (FTD), remains challenging because EEG signals are noisy, non-stationary, and vary across subjects. Segment-based learning provides a practical way to model long EEG recordings by converting them into fixed-length inputs. For each segment, discriminative information may be explored by using signals within each channel (i.e. electrode), as well as interactions between EEG channels. In this paper, we propose M-LINKX, a multi-view graph learning framework for EEG-based dementia classification. For each segment, we extract channel-level node features and construct multiple functional-connectivity (FC) graph views, where each view is defined by a specific combination of connectivity metric, frequency band, and topology filter, respectively. Instead of relying on message passing over the constructed graphs, M-LINKX follows a simple design in modeling node features and adjacency-based connectivity representations. The graph-view representations are fused using global trainable view weights, and subject-level prediction is obtained by averaging segment-level probabilities. Experiments on two three-class EEG datasets with different diagnostic groups, CAUEEG (HC/MCI/Dementia) and AHEAP (HC/AD/FTD), show that M-LINKX achieves the best subject-level performance under the main experimental settings. Our study suggests that multi-view functional connectivity can improve EEG-based dementia classification when integrated with an appropriate graph-learning architecture. Code and data are available at https://github.com/anphantt/MLINKX.

I. INTRODUCTION · II. NOTATION AND PROBLEM DEFINITION

The introduction motivates multi-view functional-connectivity graph learning for EEG-based dementia classification and presents M-LINKX as a framework that separately models channel features and graph connectivity. The notation and problem-definition section formalizes EEG channels, subject segments, training data, and classification of unseen subjects.

  • I. INTRODUCTION: Dementia-related conditions impose a growing global burden, with AD, MCI, and FTD representing clinically important diagnostic groups.More than 55 million people currently live with dementia, with projections reaching 139 million by 2050.
  • I. INTRODUCTION: EEG dementia-classification studies represent signals through end-to-end deep models or engineered temporal, spectral, and statistical descriptors.Reported approaches include CNNs, LSTM-based models, transformers, conventional machine learning, ensemble methods, and deep learning.
  • I. INTRODUCTION: Functional connectivity can be formulated as a graph whose nodes are EEG channels, edges are connectivity values, and node features preserve channel-specific information.This representation combines local channel descriptors with relational dependencies among electrode sites.
  • I. INTRODUCTION: A single FC graph may inadequately approximate neural interactions because graph-based models depend on the selected adjacency topology.Multi-view graph learning addresses this limitation by using complementary representations rather than one fixed graph.
  • I. INTRODUCTION: M-LINKX represents each EEG segment with channel-level node features and multiple FC graph views defined by connectivity metric, frequency band, and topology.The architecture is inspired by LINKX and separately models node features and adjacency information for FC-based EEG graphs.
  • I. INTRODUCTION: M-LINKX fuses graph-view representations using global trainable view weights, enabling different connectivity views to contribute unequally.The encoder separately models node features and graph-view adjacency-matrix representations.
  • II. NOTATION AND PROBLEM DEFINITION: The problem setting represents subjects through EEG segments indexed by subject and segment, with each segment containing a fixed number of time points across EEG channels.The notation defines the channel set C = {c1, . . . , cC}, where C = |C| is the number of channels, and uses Ni for the number of segments of subject Si.
  • II. NOTATION AND PROBLEM DEFINITION: Given labeled training subjects, the learning objective is to train a classifier that predicts labels for previously unseen test subjects using performance measures.The paper evaluates M-LINKX on two dementia-related EEG datasets with different diagnostic groups and baselines, reporting best subject-level performance under the main settings.

A. Graph Representation of EEG Signals

M-LINKX represents each variable-length EEG recording as fixed-length overlapping segments, then models each segment with channel-level node features and multiple functional-connectivity graph views.

  • Segmentation: Each EEG recording is divided into fixed-length overlapping segments with a sliding-window strategy to standardize inputs and make long recordings computationally manageable.The segmentation also supports subsequent feature extraction.
  • Node features: For every segment, channel-level node features are extracted into a node attribute matrix, with each row corresponding to one EEG channel.The matrix contains F node features.
  • Graph views: M-LINKX constructs K functional-connectivity graph views for each segment, representing the k-th view with an adjacency matrix.The graph views capture connectivity-based relationships among EEG channels.
  • Segment representation: Thus, each EEG segment is represented jointly by its channel-level node attributes and multiple connectivity-based graph views.The combined representation consists of X_i,t and A_i,t.

III. METHODOLOGY · A. Overall Framework · B. Node Feature Extraction

M-LINKX divides subject EEG recordings into overlapping fixed-length segments, represents each segment with channel-level node features and multiple connectivity graph views, and classifies segments through an encoder. Subject-level predictions are formed by averaging segment probabilities and selecting the highest-probability class.

  • A. Overall Framework: Each subject EEG recording is divided into fixed-length overlapping segments for segment-level graph modeling.Because labels are available only at the subject level, each segment inherits its subject’s diagnostic label during training.
  • A. Overall Framework: The framework learns an encoder that maps each segment-level graph input G_i,t to an embedding h_i,t, followed by classification into probabilities p_i,t.The classifier maps the segment embedding to a class-probability vector.
  • A. Overall Framework: Each segment supplies channel-wise node features and functional-connectivity-based graph views as inputs to the M-LINKX encoder.The encoder integrates node-feature and graph-view representations to produce a segment embedding for segment-level classification.
  • A. Overall Framework: During inference, subject-level prediction is obtained by soft voting, averaging predicted probabilities across all segments.The predicted label is the class index corresponding to the highest averaged probability.
  • B. Node Feature Extraction: For each EEG segment, channel-wise features describe local signal characteristics of individual electrodes and are concatenated within each channel.These concatenated channel features form the node feature matrix X.
  • B. Node Feature Extraction: The node feature matrix X contains one row per EEG channel and one column per selected node feature.The feature vector of the j-th EEG channel is denoted x_j ∈ R^F.

C. Functional Connectivity Extraction

For each preprocessed EEG segment, M-LINKX computes channel-pair functional connectivity for selected metric–frequency-band combinations, producing band-specific FC matrices. Coherence and wPLI capture linear coupling and consistent non-zero phase-lag relationships, respectively, and each metric–band pair forms a connectivity view with zeroed self-connections.

  • FC matrix construction: For connectivity metric m and frequency band b, pairwise EEG-channel connectivity produces a dense FC matrix W(m,b) ∈ R^C×C.Each entry represents the connectivity strength between channels c_p and c_q.
  • Connectivity metrics: The framework uses coherence and weighted phase lag index (wPLI) to construct band-specific FC matrices.Coherence measures frequency-domain linear coupling, whereas wPLI measures consistent non-zero phase-lag relationships between channels.
  • Coherence: Band-level coherence is averaged over frequencies within each band, with values ranging from 0 to 1.A value of 0 indicates no linear coupling, while 1 indicates perfect linear coupling.
  • wPLI: For each frequency band, wPLI is estimated from the sine of the instantaneous phase difference after band-pass filtering and Hilbert-transform phase extraction.wPLI values range from 0 to 1, with larger values indicating stronger and more consistent non-zero phase-lag coupling.
  • Connectivity views: The resulting connectivity matrices are symmetric with zero diagonal entries, and each selected metric–band pair becomes a graph connectivity view.Zero diagonal entries exclude self-connections, while the views are converted into graph topologies for multi-view learning.

D. Graph Construction

M-LINKX represents each EEG segment as a channel-node graph whose weighted edges are defined by functional connectivity. Graph views vary by connectivity metric, frequency band, and topology filter, with filters controlling retained edges and allowing segment-dependent topology changes.

  • Each EEG segment is represented as a graph in which EEG channels are nodes and functional-connectivity values define weighted edges.
  • Each graph view combines a connectivity metric, frequency band, and topology filter that determines which edges are retained.
  • Complete and domain topologies have fixed edge sets across segments, while top-k and hybrid topologies can vary in retained edges and edge weights.Top-k retains the strongest connectivity edges for each node, and hybrid combines domain topology with top-k edges.
  • Alpha-band coherence examples show that functional-connectivity patterns can change across adjacent segments, producing either topology changes or similar dominant connections.The examples use m = coh and b = α across consecutive segments from the same recording.

E. M-LINKX Algorithm · IV. EXPERIMENTS AND RESULTS · A. Datasets

M-LINKX separately encodes EEG node features and multiple graph-view adjacency matrices, fuses their embeddings with trainable weights, and predicts subject labels by averaging segment probabilities. It is evaluated on AHEAP and non-overlapping CAUEEG recordings using common EEG channels and overlapping segments.

  • E. M-LINKX Algorithm: Training propagates each subject label to its EEG segments, while testing averages segment-level probabilities to produce the subject-level prediction.The algorithm constructs a graph for each segment from K graph-view definitions.
  • E. M-LINKX Algorithm: M-LINKX separately encodes node features and graph-view adjacency matrices, fuses graph embeddings with trainable view weights, and classifies them with an MLP.Unlike original LINKX, M-LINKX omits additive skip connections between adjacency and node-feature embeddings.
  • IV. EXPERIMENTS AND RESULTS: The approach is evaluated on two EEG datasets for dementia-related classification, spanning different diagnostic groups and recording cohorts.AHEAP includes Alzheimer’s disease, frontotemporal dementia, and healthy control groups; CAUEEG includes dementia, mild cognitive impairment, and healthy control classes.
  • A. Datasets: CAUEEG contains EEG recordings from 1122 subjects across dementia, mild cognitive impairment, and healthy control classes, using its non-overlap version.Validation and test subjects do not overlap with the training set.
  • A. Datasets: Both datasets retain 19 common EEG channels and use sliding-window segmentation with 50% overlap to form graph instances.Segments provide inputs for node-feature extraction, functional-connectivity matrix construction, and model training.
  • A. Datasets: The main experiments use 4s segments for AHEAP and 10s segments for CAUEEG, with additional analyses across different segment lengths.The differing defaults reflect AHEAP’s 2.5 times higher sampling frequency.

B. Experimental Setup

The experiments use subject-wise evaluation protocols tailored to each dataset, with consistent splits and random seeds across models and final reporting at the subject level.

  • Evaluation protocol: AHEAP is evaluated using subject-wise 5-fold cross-validation, while CAUEEG uses the original study’s official non-overlapping train/validation/test split.CAUEEG validation and test subjects do not overlap with the training subjects.
  • Evaluation protocol: All models use the same subject-wise splits and three random seeds, with final results reported at the subject level.The passage also states that models are trained using the Adam optimizer.

C. Node-feature and Graph-view Settings

The method uses relative band power and statistical descriptors as channel-level node features, alongside four candidate functional-connectivity graph views spanning connectivity metrics, frequency bands, and topology filters.

  • Node-feature settings: Node features comprise two groups: relative band power (RBP) and statistical descriptors summarizing each EEG channel.RBP is computed from each channel’s power spectral density using Welch’s method.
  • Node-feature settings: RBP features cover delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–45 Hz).The statistical descriptors include channel-wise mean and standard deviation, with the supplied passage truncated thereafter.
  • Graph-view settings: The graph-view set combines connectivity metrics, frequency bands, and topology filters to define candidate graph representations.The experiments use four graph views.
  • Graph-view settings: The four views are (wPLI, θ, complete), (wPLI, α, domain), (coh, α, hybrid), and (coh, θ, top-k).These views span coherence and wPLI connectivity measures with theta- and alpha-band connectivity and distinct topology filters.

D. Baseline and Comparison Models · E. Performance Metrics · F. Results and Analysis

The framework is evaluated against raw-signal, feature-based, functional-connectivity, graph-based, and multi-view baselines, with subject-level balanced accuracy and macro-F1 reported for imbalanced diagnostic groups.

  • D. Baseline and Comparison Models: M-LINKX is compared with CEEDNet-style 1D-ResNet and CNN-LSTM raw-signal baselines.These comparisons assess performance against classical and deep-learning approaches operating on raw EEG signals.
  • D. Baseline and Comparison Models: An MLP trained on flattened channel-level node features provides a feature-based baseline without functional connectivity.This isolates the contribution of functional-connectivity information from channel-level node features.
  • D. Baseline and Comparison Models: Single-view GATv2 and LINKX-style models are evaluated using one functional-connectivity view at a time.Each baseline is trained separately on every candidate view, and its strongest subject-level result among the four runs is reported.
  • D. Baseline and Comparison Models: The LINKX comparison uses the same no-skip LINKX-style encoder as M-LINKX, isolating the effect of multi-view connectivity fusion.The comparison is designed to differ in whether multiple connectivity views are fused.
  • E. Performance Metrics: All reported results are evaluated at the subject level because diagnostic labels are assigned to subjects rather than individual EEG segments.Subject-level evaluation aligns the prediction unit with the diagnostic-label definition.
  • E. Performance Metrics: Balanced accuracy averages recall across classes, while macro-F1 averages class-wise F1-scores combining precision and recall.Both metrics are reported to account for imbalanced class distributions across diagnostic groups.

1) Balanced Accuracy and Macro-F1 Scores: … G. Learned Graph-view Weights

M-LINKX achieves the strongest subject-level balanced accuracy and macro-F1 across CAUEEG and AHEAP under the main settings, while its errors, segment-duration sensitivity, and learned view weights reveal dataset- and duration-dependent behavior. Performance is strongest at shorter windows, and connectivity views contribute unequally to predictions.

  • 1) Balanced Accuracy and Macro-F1 Scores:: M-LINKX achieves the best subject-level balanced accuracy and macro-F1 on both CAUEEG and AHEAP.On CAUEEG, it obtains 66.49% balanced accuracy and 66.65% macro-F1, outperforming the strongest non-proposed baseline by about 5.0% and 4.4%, respectively.
  • 1) Balanced Accuracy and Macro-F1 Scores:: M-LINKX outperforms the node-feature-only MLP, single-view LINKX, and single-view GATv2, indicating complementary and heterogeneous information from multi-view functional connectivity.The evaluated views vary by connectivity metric, frequency band, and topology filter.
  • 2) Subject Classification Confusion Matrices:: MCI is the main source of CAUEEG errors, with MCI subjects often predicted as HC and some dementia subjects predicted as MCI.The passage attributes this difficulty to shared EEG characteristics between MCI, healthy aging, and dementia.
  • 2) Subject Classification Confusion Matrices:: FTD has substantially lower recall than HC and AD in AHEAP, with most FTD errors involving predictions as HC or AD.This finding comes from normalized subject-level confusion matrices.
  • 3) Algorithm Performance vs. Segment Lengths:: M-LINKX remains competitive across segment durations, but its advantage over baselines becomes less consistent as window length increases.The passage states that longer windows may stabilize feature-based baselines without necessarily improving multi-view connectivity learning.
  • 3) Algorithm Performance vs. Segment Lengths:: 10-second windows yield M-LINKX’s best CAUEEG performance, after which its performance decreases while several baselines remain relatively stable.Longer windows may smooth transient connectivity patterns useful to the multi-view graph representation.
  • 3) Algorithm Performance vs. Segment Lengths:: On AHEAP, some baselines improve with longer segments, whereas M-LINKX shows a decreasing performance trend.Longer windows may stabilize node-feature or single-view representations while making multiview connectivity graphs less distinctive.
  • G. Learned Graph-view Weights: Learned graph-view weights are nonuniform, showing that different metric-band-topology views contribute differently across datasets and segment lengths.On CAUEEG, coherence-alpha hybrid dominates at 10 seconds, while weights become more balanced at 20 seconds; on AHEAP, wPLI-alpha domain dominates at 2 seconds and coherence-theta top-k at 4 and 6 seconds.

H. Discussion · V. CONCLUSION AND FUTURE WORK

The discussion attributes M-LINKX’s performance to appropriately modeled multi-view functional connectivity, while noting segment-length dependence and difficulty separating clinically related diagnostic groups. The conclusion presents M-LINKX’s multi-view graph-learning design and identifies adaptive graph-view selection as future work.

  • H. Discussion: Multi-view functional-connectivity learning can improve EEG segment-based dementia classification when connectivity information is modeled appropriately.
  • H. Discussion: Learned graph-view weights are relatively stable across segment lengths but nonuniform, indicating unequal use of views during fusion.The weights describe model behavior and should not be interpreted as direct biological importance.
  • H. Discussion: The contribution of connectivity-based models varies across segment lengths, whose optimal duration may depend on dataset, disease groups, and model architecture.Functional-connectivity estimates and retained graph topologies can change with EEG window duration.
  • H. Discussion: Some diagnostic groups remain difficult to separate, especially clinically related classes such as HC/MCI and AD/FTD.
  • V. CONCLUSION AND FUTURE WORK: M-LINKX is proposed as a multi-view graph learning framework for EEG-based dementia classification, using channel correlations as potentially useful information for subject classification.
  • V. CONCLUSION AND FUTURE WORK: Each EEG segment combines channel-level node features with multiple functional-connectivity graph views formed from different connectivity metrics, frequency bands, and topology filters.
  • V. CONCLUSION AND FUTURE WORK: M-LINKX uses a simple heterophilic graph-learning backbone that treats node features and adjacency matrices as separate information sources.
  • V. CONCLUSION AND FUTURE WORK: Future work will explore adaptive graph-view selection that assigns different importance to connectivity views across datasets, subjects, or disease groups.
Loading 2608.14847v1…