ELPG-DTFS: Adaptive GNN for EEG Depression Diagnosis

• The paper introduces a novel prior-guided adaptive time-frequency graph neural network that overcomes static time-frequency modeling, fixed connectivity, and lack of neuroscience priors in EEG depression diagnosis.
• It employs channel-band attention, cross-band mutual information, and a learnable adjacency matrix to capture dynamic functional connectivity and temporal evolution in 128-channel EEG recordings.
• The integration of residual knowledge-graph fusion with multi-scale priors boosts interpretability and robustness, achieving superior accuracy (97.63%) and F1 (97.33%) on the MODMA dataset.

Searching arXiv for ELPG-DTFS and closely related EEG depression diagnosis work. ELPG-DTFS is a Prior-Guided Adaptive Time-Frequency Graph Neural Network for EEG Depression Diagnosis proposed for objective, accurate, and interpretable EEG-based screening of major depressive disorder (MDD) (Qiu et al., 29 Sep 2025). The model is designed around three stated limitations of earlier EEG depression diagnosis methods: static time-frequency modeling, rigid or fixed connectivity graphs, and the lack of explicit neuroscience priors. Its pipeline combines channel-band attention with cross-band mutual information, a learnable adjacency matrix for adaptive functional connectivity, and a residual knowledge-graph pathway that injects prior knowledge into graph reasoning. On the 128-channel MODMA dataset (53 subjects), the reported performance is 97.63% accuracy and 97.33% F1, exceeding the reported 2025 state of the art, ACM-GNN (Qiu et al., 29 Sep 2025).

1. Clinical setting and problem formulation

ELPG-DTFS is situated in the context of EEG-based screening for MDD, where the motivating concern is that diagnosis still relies on self-rating / peer-rating scales, clinician experience, and subjective judgment (Qiu et al., 29 Sep 2025). The paper frames EEG as suitable for this application because it is non-invasive, low-cost, portable, and has high-temporal-resolution. It further states that MDD affects more than 350 million people, which establishes the clinical relevance of timely screening.

The system objective is not framed as general EEG decoding; it is specifically aimed at depression diagnosis from resting-state EEG. The paper argues that prior approaches are limited in three precise ways. First, many CNN-based methods treat spectrograms or EEG time-frequency maps as fixed images, which can blur transient neural bursts and short-lived affective dynamics. Second, graph-based models often depend on fixed or binarized Pearson-correlation graphs, which may discard weak but clinically meaningful couplings, fail to reflect time-varying connectivity, and represent edge structure too coarsely. Third, prior models often learn solely from data and therefore ignore explicit prior knowledge such as region-specific roles, hemispheric symmetry, distance-based attenuation, and clinically relevant brain partitions.

Within that framing, ELPG-DTFS is defined by three architectural additions: channel-band attention with cross-band mutual information, adaptive graph construction via a learnable adjacency matrix, and a residual knowledge-graph / prior fusion pathway. This suggests that the model is intended not only to raise classification accuracy, but also to increase structural fidelity to neurophysiological organization and to improve interpretability under small-sample EEG conditions.

2. Time-frequency representation and node feature construction

The input pipeline begins with 5-minute eyes-closed resting-state recordings from the MODMA dataset, sampled at 250 Hz (Qiu et al., 29 Sep 2025). The preprocessing sequence is specified as follows: discard the first and last 10 seconds, keep 280 seconds, apply 0.3–30 Hz FIR filtering, perform baseline subtraction, carry out ICA artifact removal, apply electrode-wise $\ell_2$ normalization, and then segment the signals into 4-second windows with 50% overlap. Each 280-second recording is thereby divided into $T = 139$ epochs per subject.

Feature construction combines bandwise differential entropy and cross-band dependence. For each window, ELPG-DTFS extracts Differential entropy (DE) in four bands, namely $\delta$, $\theta$, $\alpha$, and $\beta$, and also computes Mutual information (MI) between band pairs. The paper defines mutual information as

$\mathrm{MI}(b_1,b_2)=H(b_1)+H(b_2)-H(b_1,b_2),$

and states that the DE of a zero-mean narrow-band process is

$h(x)=\frac12 \log(2\pi e \sigma^2),$

where $\sigma^2$ is the band power.

These features are then subjected to a factorized channel-band attention mask. Let $\mathbf{a}_{\text{chan}} \in [0,1]^N$ and $T = 139$0. The attention mask is

$T = 139$1

and the DE tensor is reweighted as

$T = 139$2

The paper further interprets this mechanism as a form of variational feature selection through a binary mask $T = 139$3 with objective

$T = 139$4

Here, $T = 139$5 is a learned Bernoulli distribution over selected channel-band features, $T = 139$6 is a sparsity-promoting prior, and $T = 139$7 is a regularization weight.

After concatenating the refined DE features and MI features across time,

$T = 139$8

the sequence is fed to a BiLSTM with hidden size 64, producing node embeddings

$T = 139$9

The stated purpose is to preserve temporal evolution across windows rather than treating EEG as static.

3. Adaptive graph construction and prior-guided multi-scale reasoning

A central element of ELPG-DTFS is that inter-channel connectivity is not fixed. The graph starts from a seed adjacency constructed from Pearson correlations,

$\delta$0

and introduces a trainable mask

$\delta$1

so that the final adjacency becomes

$\delta$2

The paper states that $\delta$3 is initialized to 0.5, allowing edges to be strengthened or weakened during training. This adaptive mechanism is described as enabling the model to downweight unhelpful edges, preserve weak but meaningful interactions, and adapt connectivity rather than relying on a rigid fixed graph (Qiu et al., 29 Sep 2025).

The prior-guided component is explicitly multi-scale. At the local level, the model incorporates a distance-based affinity derived from electrode spacing:

$\delta$4

This local prior is added to the learned graph before normalization and encodes the biophysical notion that nearby electrodes are more likely to exhibit coupled activity.

At the mesoscopic level, the brain is partitioned into nine cortical groups based on the Desikan atlas and hemispheric symmetry. For each group $\delta$5, the model learns a virtual node

$\delta$6

The paper describes this as a self-attention pooling mechanism that summarizes local regional dynamics while keeping the graph compact.

At the global level, 3D electrode positions $\delta$7 are embedded through sinusoidal encoding,

$\delta$8

followed by a six-head attention layer that constructs a global adjacency,

$\delta$9

To limit graph density, only the top 25% edges per node are retained. Downstream graph processing uses two spectral graph convolution layers,

$\theta$0

with normalized Laplacian

$\theta$1

and ReLU nonlinearity $\theta$2.

The paper’s prior pathway is not expressed as a separate explicit knowledge-graph loss. Instead, it is implemented as gated residual fusion between data-driven embeddings and prior-encoded features:

$\theta$3

Here $\theta$4 denotes prior knowledge encoding. The gate is explicitly initialized so that its norm is small at the start, which means that learning begins mostly data-driven and gradually incorporates prior information. The priors named in the paper include brain-region roles, structural links, clinical heuristics, and local/mesoscopic/global graph priors.

4. Training protocol and evaluation methodology

The reported dataset comprises 53 subjects total, specifically 24 MDD patients and 29 healthy controls, all drawn from MODMA (Qiu et al., 29 Sep 2025). Validation uses subject-wise 10-fold cross-validation, with no subject leakage between train and test, and 10% of each training fold used as validation. This evaluation design is important because the paper’s claims are based on subject-level rather than epoch-level separation.

Optimization is implemented in PyTorch with the Adam optimizer, learning rate $\theta$5, weight decay $\theta$6, and batch size 32. The training regime uses early stopping after 10 epochs without validation improvement, and the experiments are run on an RTX 3090 GPU. The paper also states that training stabilizes within about 80 epochs.

The reported evaluation metrics are Accuracy (Acc), Precision (Pre), Recall (Rec), and F1-score (F1), all presented as mean ± standard deviation across 10 folds. Statistical significance is tested using the Wilcoxon signed-rank test with $\theta$7. This suggests that the empirical claims are framed not only in terms of raw point estimates but also in terms of fold-wise variability and nonparametric significance testing.

A potential misconception is that ELPG-DTFS is evaluated as a purely black-box classifier on static EEG images. The described protocol contradicts that interpretation: the model explicitly uses sliding-window temporal modeling, adaptive graph learning, and prior-guided graph reasoning. Another possible misconception is that the “knowledge-graph” component implies a separate symbolic reasoning module. The paper instead describes a residual knowledge-graph / prior fusion pathway, with prior injection realized directly in the learned representation space.

5. Reported performance and ablation evidence

On MODMA, ELPG-DTFS achieves Accuracy: 97.63%, Precision: 96.68%, Recall: 98.03%, and F1-score: 97.33% (Qiu et al., 29 Sep 2025). The paper emphasizes that these results surpass ACM-GNN, which is reported at 95.46% Accuracy and 95.80% F1. The stated gains are +2.17 percentage points Accuracy and +1.53 percentage points F1, with significance reported as $\theta$8. The paper also states that ELPG-DTFS surpasses the 2025 state of the art by up to 2.57 percentage points in performance comparison discussions.

The ablation results are central to the model’s claimed rationale. For the full model, the paper reports Acc: 97.63 and F1: 97.33. Removing prior knowledge reduces performance to Acc: 95.12 and F1: 95.29. Removing the learnable adjacency gives Acc: 94.01 and F1: 94.58. Removing MI gives Acc: 94.85 and F1: 95.03. Removing attention + MI produces the largest degradation, namely Acc: 93.25 and F1: 92.98, corresponding to an F1 drop of 4.35 points.

The paper interprets these ablations in component-specific terms. The prior pathway is said to improve both precision and recall, suggesting fewer false positives and better detection of true MDD cases. The learnable adjacency is described as crucial for detecting subtle pathological connectivity, and the paper notes that recall drops most when this module is removed. MI is said to capture nonlinear cross-band dependencies and help suppress confusion with healthy spectral variation. The joint removal of attention and MI produces the largest decline, which the paper interprets as evidence that channel-band attention and cross-band MI are the most important contributors to discriminative power.

These results do not establish universal performance across all EEG depression datasets; they establish the reported behavior on MODMA under the stated cross-validation protocol. A plausible implication is that the model’s gains are closely linked to the combined effect of feature selection, adaptive connectivity, and prior injection rather than to any single architectural modification in isolation.

6. Interpretability, methodological significance, and scope

The paper presents ELPG-DTFS as interpretable for four stated reasons (Qiu et al., 29 Sep 2025). First, channel-band attention weights indicate which electrodes and rhythms are emphasized. Second, the learned adjacency matrix exposes which inter-channel connections are treated as important. Third, the prior-guided graph pathway aligns learned behavior with neuroscience knowledge. Fourth, the multi-scale structure supports interpretation at the local electrode level, mesoscopic brain-region level, and global connectivity level.

Methodologically, ELPG-DTFS is positioned against models that treat spectra as static images and graphs as fixed structures. In contrast, it preserves temporal evolution through the BiLSTM, models cross-frequency dependence via MI, adapts functional connectivity through a trainable adjacency mask, and introduces neuroscience priors through local distance affinity, mesoscopic virtual centers, global transformer adjacency, and gated residual prior fusion. This suggests an overview of time-frequency analysis, adaptive graph learning, and prior-guided representation learning within a single EEG classification framework.

The model’s relation to ACM-GNN is presented as both empirical and architectural. Empirically, ELPG-DTFS reports higher Accuracy and F1 on MODMA. Architecturally, the paper attributes the difference to the addition of explicit channel-band attention, cross-band mutual information, a learnable adjacency mask, multi-scale priors, and gated residual prior fusion. In that sense, ELPG-DTFS is not described merely as a deeper or larger graph network, but as a graph framework that combines dynamic connectivity estimation with structured neurophysiological priors.

The scope of the reported evidence remains bounded by the experimental setting in the paper: 128-channel EEG, 53 subjects, and subject-wise 10-fold cross-validation on MODMA. Within that scope, ELPG-DTFS is presented as a robust and interpretable framework for EEG-based MDD diagnostics. A plausible implication is that its primary technical contribution lies in showing how adaptive graphs and multi-scale priors can be integrated into time-frequency EEG modeling without reducing the problem to static image classification.