Bridging Accuracy and Explainability in EEG-based Graph Attention Network for Depression Detection

TL;DR

This work tackles the problem of accurately screening major depressive disorder from EEG while ensuring interpretability. It introduces ExPANet, a graph neural network that models EEG data as brain graphs with 19 electrode nodes and PLV-based edges, enriched by 14 multi-domain features per node and a graph attention mechanism with edge gates, AxisMix fusion, and a virtual node for global context. The approach achieves state-of-the-art performance on two datasets (Dataset-I: $97.5\%$ accuracy; Dataset-II: $91.4\%$ accuracy) and provides multi-level explanations, identifying key features (nonlinear/fractal metrics), critical channels (frontal/central), and connectivity patterns (reduced occipito-frontal/occipito-temporal links in MDD) that align with clinical findings. This combination of high accuracy and mechanistic interpretability supports reliable, clinically relevant EEG-based screening for depression and demonstrates a path toward explainable AI in mental health diagnostics.

Abstract

Depression is a major cause of global mental illness and significantly influences suicide rates. Timely and accurate diagnosis is essential for effective intervention. Electroencephalography (EEG) provides a non-invasive and accessible method for examining cerebral activity and identifying disease-associated patterns. We propose a novel graph-based deep learning framework, named Edge-gated, axis-mixed Pooling Attention Network (ExPANet), for differentiating major depressive disorder (MDD) patients from healthy controls (HC). EEG recordings undergo preprocessing to eliminate artifacts and are segmented into short periods of activity. We extract 14 features from each segment, which include time, frequency, fractal, and complexity domains. Electrodes are represented as nodes, whereas edges are determined by the phase-locking value (PLV) to represent functional connectivity. The generated brain graphs are examined utilizing an adapted graph attention network. This architecture acquires both localized electrode characteristics and comprehensive functional connectivity patterns. The proposed framework attains superior performance relative to current EEG-based approaches across two different datasets. A fundamental advantage of our methodology is its explainability. We evaluated the significance of features, channels, and edges, in addition to intrinsic attention weights. These studies highlight features, cerebral areas, and connectivity associations that are especially relevant to MDD, many of which correspond with clinical data. Our findings demonstrate a reliable and transparent method for EEG-based screening of MDD, using deep learning with clinically relevant results.

Figures (7)

• Figure 1: Illustration of the proposed ExPANet architecture for EEG-based classification of MDD vs HC.
• Figure 2: Evaluation of model performance on both datasets with error bars for all metrics. Round markers represent Dataset-I and square markers represent Dataset-II. The error bars represent inter-fold variability. The plot shows the superior consistent performance of the proposed model in comparison to current baselines across both datasets.
• Figure 3: Group-wise node-feature importance derived from GNNExplainer and averaged within HC vs MDD for each dataset. Bars are ordered low to high, to highlight the most discriminative features per group.
• Figure 4: Channel importance rankings for HC and MDD across both datasets. Fronto–central electrodes consistently rank higher than posterior sites.
• Figure 5: Analysis of connectivity importance. Only edges consistently weaker in MDD than HC are shown, with a color scale indicating the degree of reduction (higher values indicate stronger weakening). Both datasets show weakened long-range occipito-temporal and occipito-frontal connections in MDD, although residual fronto-midline couplings remain dominant.