Early Mild Cognitive Impairment Diagnosis via Resting-State fMRI Brain Networks Using a Region-Specific Hierarchical Fusion Graph Neural Network

Abstract

Early mild cognitive impairment (EMCI) is the earliest intervenable stage of Alzheimer’s disease (AD). Although graph neural networks (GNNs) have begun to exploit brain network topology, traditional fMRI-based diagnostic methods often neglect these structural patterns by relying on vectorized features. Furthermore, existing GNNs frequently disregard inter-regional functional heterogeneity and group-level discriminative patterns, leading to limited accuracy and biomarker interpretability. To address these challenges, we propose HF-BrainGNN, an end-to-end hierarchical graph learning framework for EMCI identification. Our method introduces a functional affinity region convolution (FAR-Conv) layer to learn region-adaptive kernels, a Differential Focus Pooling (DF-Pool) module to identify disease-salient brain regions by maximizing inter-group distinctiveness, and a hierarchical integration classifier (HIC) to fuse multi-level graph representations. The framework is optimized using classification, focus separation, and consistency regularization losses. Experiments on the ADNI dataset (104 EMCI, 114 Cognitively Normal) show that HF-BrainGNN achieves 86.78% accuracy, outperforming the best baseline (Hi-GCN) by 4.64%. Furthermore, the automatically identified regions, such as the bilateral hippocampus and default mode network hubs, align with established EMCI biomarkers. Ultimately, HF-BrainGNN provides an efficient, interpretable artificial intelligence tool for precise brain network characterization and early AD intervention.

1. Introduction

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder affecting more than 50 million individuals worldwide, and the number is projected to reach 152 million by 2050 [1]. AD typically develops from mild cognitive impairment (MCI), whereas early mild cognitive impairment (EMCI) represents the earliest clinically detectable and potentially intervenable stage [2]. Although EMCI manifests only a subtle cognitive decline that does not severely impact daily life, its timely and accurate identification is crucial for delaying the progression toward dementia [3]. Due to the mild and non-specific nature of EMCI symptoms, diagnosis based solely on clinical evaluation remains challenging, highlighting the need for objective neuroimaging-based biomarkers [4].

Functional magnetic resonance imaging (fMRI) provides a non-invasive means of quantifying brain functional activity and connectivity, offering critical insight into early neural abnormalities associated with AD [5]. Resting-state fMRI (rs-fMRI), which requires no cognitive task, is particularly suitable for clinical and research settings involving patients with varying levels of cognitive ability. The brain functional connectivity networks derived from rs-fMRI reflect both local and global neural interactions and have been widely used to investigate the full AD continuum [6,7].

Traditional approaches for rs-fMRI analysis commonly rely on seed-based correlation mapping, independent component analysis (ICA), or graph-theoretic metrics [8,9]. These methods typically follow a two-stage pipeline: handcrafted feature extraction followed by classification via conventional machine-learning algorithms. For instance, Khazaee et al. [10] employed graph-theoretic features combined with support vector machines (SVMs) for AD classification; Wee et al. [11] integrated volumetric and diffusion tensor imaging (DTI) data in multimodal models; and Tijms et al. [12] summarized graph-based biomarkers across the AD spectrum. Although informative, such two-stage strategies depend heavily on manual feature design and prior knowledge, limiting end-to-end optimization and interpretability [13].

Recently, deep learning—especially graph neural networks (GNNs)—has offered a powerful framework for modeling brain connectivity data [14]. Through message passing on graph structures, GNNs jointly learn node attributes and topological relations in an end-to-end manner, enabling automated discovery of discriminative connectivity patterns. Notable examples include population-graph GNNs for autism spectrum disorder [15], the BrainGNN architecture for fMRI-based biomarker detection [16], and dynamic attention networks for capturing temporal connectivity variations [17]. These advances demonstrate that GNNs can extract complex, informative representations of brain disorders.

However, several critical challenges remain in applying GNNs to fMRI-based cognitive impairment analysis. First, most models adopt unified convolution kernels for all nodes, disregarding the functional specialization and anatomical diversity among brain regions. Second, existing graph pooling methods predominantly assess node importance within individual graphs, neglecting systematic differences between subject groups—thereby limiting their ability to identify disease-specific regions. Third, many GNN frameworks depend only on final-layer features for classification, ignoring useful multi-level representations obtained during hierarchical propagation. Incorporating multi-hop neighborhood information across layers would allow more comprehensive integration of local and global connectivity patterns.

Specifically, to address regional heterogeneity, we introduce the functional affinity region convolution (FAR-Conv) layer. It constructs region-specific kernels by integrating functional similarity with anatomical proximity, enabling the model to capture diverse inter-regional dependencies. Simultaneously, to uncover group-level biomarkers, the differential focus pooling (DF-Pool) layer compares inter-group focus score distributions to identify disease-relevant regions, performing dimensionality reduction via a soft mask-based mechanism. Finally, a hierarchical integration classifier (HIC) is designed to adaptively fuse multi-level features with learnable weights, ensuring a robust diagnosis by combining local and global perspectives.

Furthermore, the framework optimizes three complementary loss functions—classification, focus separation, and consistency regularization—to strike a balance between diagnostic accuracy and biological interpretability.

The proposed HF-BrainGNN was evaluated on resting-state fMRI data from the ADNI cohort, including 104 EMCI patients and 114 cognitively normal individuals. Experimental results demonstrate superior classification performance compared with existing methods, achieving improved accuracy, sensitivity, and specificity. Moreover, the key brain regions identified by DF-Pool correspond closely to EMCI-associated areas reported in previous neuroscience studies [18], underscoring the physiological plausibility and clinical interpretability of the proposed model.

2. Method

This section elaborates on the HF-BrainGNN framework for EMCI classification. As illustrated in Figure 1, the overall workflow consists of three distinct stages: (1) brain network construction (Section 2.1), where fMRI data is processed via atlas-based parcellation to generate functional connectivity networks; (2) HF-BrainGNN learning, which extracts hierarchical features and identifies discriminative regions (Section 2.2, Section 2.3, Section 2.4, Section 2.5 and Section 2.6); and (3) diagnosis and interpretation, where the model outputs classification results and highlights potential biomarkers. The following subsections detail each component of this pipeline.

2.1. Brain Network Construction

In this study, the Automated Anatomical Labeling (AAL90) template is used to parcellate the brain into N anatomical regions ($N=90$), providing a standardized anatomy-based reference widely adopted in neuroimaging studies [19]. Each region of interest (ROI) corresponds to one node in the brain graph; namely, $V=\{v_1,v_2,\dots ,v_N\}$.

For each subject, a Pearson correlation matrix $R\in {\mathbb{R}}^{N\times N}$ is first computed based on the mean time series of all ROIs. Each element $R_{ij}$ reflects the functional connectivity (FC) strength between regions $v_i$ and $v_j$. The adjacency matrix A is then derived from R to define the graph topology. To emphasize the most significant functional connections, we threshold the correlation matrix by retaining only the top 20% of absolute correlation coefficients. Specifically, elements within the top 20th percentile are set to 1 (indicating an existing edge), and the remainder are set to 0, thereby forming a binary adjacency matrix A.

After constructing the brain graph $G=\{V,A\}$, node features are defined from the same FC matrix. The feature of node $v_i$ is represented by its corresponding FC vector ${\mathbf{x}}_i=R_i$, where $R_i$ denotes the i-th row of R. Thus, the initial node feature matrix can be expressed as ${\mathbf{X}}^{\left(0\right)}={[{\mathbf{x}}_1,{\mathbf{x}}_2,\dots ,{\mathbf{x}}_N]}^T$. In this study, only FC information is used to form node features, without introducing additional morphological or clinical covariates [20]. All notations used in this section are summarized in

Table 1. Notations and their descriptions.

Notation	Description
N	Number of brain regions (nodes)
$v_i$	i-th node (brain region) in the graph
${\mathcal{N}}^{\left(l\right)}\left(i\right)$	Neighbor set of node $v_i$ at the l-th layer
A	Adjacency matrix, $A\in {\mathbb{R}}^{N\times N}$
$\mathbf{X}$	Node feature matrix
${\mathbf{x}}_i^{\left(l\right)}$	Feature representation of node $v_i$ at the l-th layer
$H^{\left(l\right)},{\mathbf{h}}_i^{\left(l\right)}$	Hidden feature representation output by FAR-Conv at the l-th layer
$W_i^{\left(l\right)}$	Graph convolution kernel of node $v_i$ at the l-th layer
$W_j^{\left(l\right)}$	Graph convolution kernel of neighbors of node $v_i$
${\alpha }_i$	Regional affinity vector of node $v_i$
$e_{ij}$	Edge weight enhancement factor based on regional affinity
${\mathbf{n}}_K$	Set of brain regions in the K-th functional network
$\mathrm{pos}\left(v_i\right)$	3D coordinates of brain region $v_i$
$C_K$	Centroid coordinates of network K
$s^{\left(l\right)}$	Node focus score at the l-th layer
${\tilde{s}}^{\left(l\right)}$	Normalized node focus score
${\lambda }_1,{\lambda }_2$	Hyperparameters of the loss function

.

2.2. Overview of the HF-BrainGNN Method

The proposed hierarchical focused brain graph neural network (HF-BrainGNN) is illustrated in Figure 2a. It comprises three main components: the FAR-Conv layer, the differential focus pooling layer, and the hierarchical integration classifier. The model takes a single-subject brain graph $G=(V,A)$ and its initial node features ${\mathbf{X}}^{\left(0\right)}$ (formed by FC row vectors) as input.

As shown in Figure 2a, HF-BrainGNN contains two cascaded processing layers, each integrating a FAR-Conv layer, a differential focus pooling (DF-Pool) layer, and a global pooling operation (Mean/Max Pool) to obtain layer-specific graph representations (${\mathbf{g}}_1$, ${\mathbf{g}}_2$). These representations are adaptively fused in the hierarchical integration module and fed into a multilayer perceptron (MLP) for EMCI/CN classification.

Within each hierarchical brain processing unit (HBPU), FAR-Conv computes regional affinity vectors to generate region-specific convolution kernels and performs two-hop message passing, enabling multi-level neighborhood aggregation to yield representations ${\mathbf{H}}^{\left(l\right)}$ for the two cascaded processing layers ($l=1,2$). DF-Pool then calculates inter-group-enhanced focus scores to select the top-k disease-relevant brain regions and outputs the induced subgraph for subsequent layer-wise processing.

Finally, the hierarchical integration classifier performs weighted fusion of the readout vectors from all layers and outputs the final class probabilities through the MLP. The entire pipeline proceeds from feature extraction (FAR-Conv) to region selection (DF-Pool), and then to feature integration and classification. Three complementary loss functions are jointly optimized to ensure both strong predictive performance and interpretability, as detailed in Section 2.6.

2.3. Functional Affinity Region Convolutional Layer

Graph neural networks have shown promising potential in brain network analysis; however, there is a notable limitation in their application: existing models typically use a unified convolution kernel for feature extraction across all nodes. This approach may fail to fully capture the unique functional characteristics and anatomical positional differences of distinct brain regions.

Neuroscience studies have demonstrated that brain regions exhibit clear functional specialization and complex interaction patterns [21]. For instance, the prefrontal cortex primarily mediates executive functions, while the temporal lobe is involved in auditory and language processing. This functional differentiation and regional specificity are critical for a comprehensive understanding of brain network dynamics, and thus require special consideration in computational model design.

To address this issue, we propose the FAR-Conv layer, which accounts for the functional similarity and anatomical positional characteristics of brain regions, assigning region-specific embedding strategies to different brain areas. The core idea of FAR-Conv is to enhance information exchange between functionally similar brain regions while preserving the anatomical specificity of each region, as illustrated in Figure 2b. For each brain region node $v_i$, we introduce a regional affinity vector ${\alpha }_i\in {\mathbb{R}}^{K+1}$, which represents the membership degree of node $v_i$ to each functional region [22]. In this study, $K=8$ functional networks are defined according to the commonly adopted eight-network division widely used in neuroscience research [23]. To incorporate multi-dimensional positional information $p_i$, the regional affinity vector is defined as ${\alpha }_i=[d_{i1},d_{i2},\dots ,d_{iK},p_i]$, where $d_{ik}={\parallel c_i-C_k\parallel }_2$ denotes the Euclidean distance from brain region i to the centroid of the k-th functional network; $c_i\in {\mathbb{R}}^3$ represents the normalized 3D coordinate of the brain region’s center, $p_i$ is the positional encoding of node $v_i$.

The positional encoding adopts the following form:

$$p_i=\left[sin\left({\displaystyle \frac{c_i}{\tau }}\right),cos\left({\displaystyle \frac{c_i}{\tau }}\right)\right]$$

(1)

where $c_i=(x_i,y_i,z_i)$ is the 3D coordinate vector of the brain region, and $\tau$ is a scale parameter. This formulation provides a 6-dimensional positional embedding (3 sine and 3 cosine components), enabling the model to distinguish brain regions located in different anatomical positions even if they exhibit similar FC patterns.

The centroid coordinate of the k-th functional network is calculated as follows:

$$C_K={\displaystyle \frac{1}{|{\mathbf{n}}_K|}}\sum _{v_i\in {\mathbf{n}}_K}\mathrm{pos}\left(v_i\right)$$

(2)

where $C_K$ is the centroid coordinate of network K, ${\mathbf{n}}_K$ is the set of brain regions in the k-th functional network, and $\mathrm{pos}\left(v_i\right)$ denotes the 3D coordinate of brain region $v_i$.

Based on the regional affinity, the region-specific convolution kernel $W_i^{\left(l\right)}$ for each node is computed as follows:

$$W_i^{\left(l\right)}=W_{\mathrm{share}}^{\left(l\right)}+{\displaystyle \sum _{k=1}^K}\mathrm{softmax}(-d_{ik})B_k^{\left(l\right)}$$

(3)

where $W_{\mathrm{share}}^{\left(l\right)}$ is the shared base weight, and $B_k^{\left(l\right)}$ is the base weight matrix of the k-th functional network. Here, $d_{ik}$ represents the Euclidean distance between brain region $v_i$ and the centroid of network k. To ensure that brain regions located closer to a functional network receive higher aggregation weights, we apply the negative distance ($-d_{ik}$) within the softmax function, effectively converting the distance metric into a similarity measure. The superscript l indexes the hierarchical processing layer, with each layer comprising one FAR-Conv and one DF-Pool operation. Two such cascaded layers are stacked in the proposed architecture.

Information transfer between functionally similar brain regions is realized via an edge weight enhancement factor:

$$e_{ij}=e_{ij}^0·\sigma \left({\alpha }_i^T{\alpha }_j\right)$$

(4)

Here, $\sigma$ denotes the sigmoid function, which ensures the enhancement factor is constrained to the range $(0,1)$, and $e_{ij}^0$ is the original edge weight. This design strengthens message passing between node pairs with similar regional affinity, while weakening message passing between dissimilar node pairs.

The final forward propagation of FAR-Conv is given by the following:

$${\mathbf{h}}_i^{\left(l\right)}=\mathrm{ReLU}\left(W_i^{\left(l\right)}{\mathbf{x}}_i^{\left(l\right)}+\sum _{j\in {\mathcal{N}}^{\left(l\right)}\left(i\right)}{e}_{ij}·W_j^{\left(l\right)}{\mathbf{x}}_j^{\left(l\right)}\right)$$

(5)

where ${\mathbf{x}}_i^{\left(l\right)}$ represents the feature representation of node $v_i$ at the l-th layer, $W_i^{\left(l\right)}$ is the region-specific graph convolution kernel of node $v_i$, $e_{ij}$ is the edge weight enhancement factor based on regional affinity, ${\mathcal{N}}^{\left(l\right)}\left(i\right)$ denotes the neighbor set of node $v_i$ at the l-th layer, and $W_j^{\left(l\right)}$ is the region-specific graph convolution kernel of the neighbor node $v_j$.

2.4. Differential Focus Pooling

Graph pooling, a fundamental operation in graph neural networks, plays a crucial role in reducing computational complexity and extracting salient features [24,25]. However, conventional pooling methods mainly evaluate node importance within individual graphs or datasets, without explicitly modeling inter-group discriminative patterns. In brain network analysis, this limitation reduces the ability to identify disease-specific key regions [26].

Motivated by these observations, we propose the differential focus pooling (DF-Pool) mechanism (Figure 3), which automatically identifies disease-related brain regions by comparing node-level representations across subject groups: it first computes node focus scores for each subject, then estimates inter-group differences to highlight highly discriminative regions, and finally performs threshold-based pooling to retain the most informative nodes, enabling targeted feature extraction and effective dimensionality reduction. Grounded in a widely accepted neuroimaging principle—that brain regions with high diagnostic value typically exhibit significant group-wise differences between patients and healthy controls [27] (a principle validated across multiple neuropsychiatric disorders and providing a theoretical basis for discovering informative differential biomarkers)—we design a three-step procedure for DF-Pool implementation: (1) node focus score calculation, (2) inter-group difference estimation, and (3) differential node selection.

First, we compute the focus score for each node in a brain graph to quantify its contribution to the classification objective:

$$f_m={\mathbf{H}}_m^{\left(l\right)}·{\mathbf{v}}^{\left(l\right)}/{\parallel {\mathbf{v}}^{\left(l\right)}\parallel }_2,$$

(6)

where ${\mathbf{H}}_m^{\left(l\right)}$ is the hidden feature matrix output by FAR-Conv at the l-th layer, $v^{\left(l\right)}$ is a learnable projection vector, and $\parallel v^{\left(l\right)}{\parallel }_2$ is its normalization factor. The focus score $f_m$ measures the projection strength of node features along the direction of $v^{\left(l\right)}$, reflecting node saliency with respect to the learned representation.

We then apply min–max normalization to constrain the focus scores within the interval $[0,1]$:

$${\tilde{f}}_m=\left(f_m-min\left(f_m\right)\right)/\left(max\left(f_m\right)-min\left(f_m\right)\right),$$

(7)

which ensures a consistent range and preserves the relative ranking of node importance, preventing bias from extreme values.

Next, we estimate inter-group differences among subjects. For each class c, we compute its group-level average focus score $F_c$ (Equation (8)), and derive the inter-group difference matrix D (Equation (9)):

$$F_c={\displaystyle \frac{1}{|S_c|}}\sum _{m\in S_c}{\tilde{f}}_m,$$

(8)

$$D=|F_{c_1}-F_{c_2}|\odot {\displaystyle \frac{F_{c_1}+F_{c_2}}{2}},$$

(9)

where $S_c$ denotes the set of subjects belonging to class c, and $|S_c|$ is the number of subjects. The Hadamard product operator (⊙) jointly encodes two complementary factors [28]: the absolute difference $|F_{c_1}-F_{c_2}|$ quantifies the inter-group discrepancy, while the average activation level $(F_{c_1}+F_{c_2})/2$ captures the overall node activity. This design emphasizes brain regions that show both strong group-wise discriminability and high activation, which are likely to correspond to disease-specific biomarkers. The group-level average focus $F_c$ and difference matrix D serve as intermediate representations that guide the subsequent node selection, emphasizing nodes with stronger inter-group discriminability.

Finally, threshold-based pooling is performed to retain the most informative nodes:

$${\mathbf{s}}_m^{\mathrm{pool}}={\tilde{f}}_m\odot D,$$

(10)

During the implementation, the group-level average focus $F_c$ and the difference matrix D are computed solely using the subjects in the training set to prevent data leakage. In the inference phase, the matrix D remains fixed as a learned prior. For a single unseen subject, the pooling operation (Equation (10)) relies only on the subject’s individual focus score and the pre-trained D, ensuring the diagnosis is independent of ground-truth labels or other test data.

$$\begin{array}{cc}\hfill {\mathbf{X}}_m^{(l+1)}& ={({\mathbf{H}}_m^{\left(l\right)}\odot {\tilde{f}}_m)}_{idx},\hfill \end{array}$$

(11)

$$\begin{array}{cc}\hfill A_m^{(l+1)}& =A_m^{\left(l\right)}(idx,idx),\hfill \end{array}$$

(12)

where $idx$ denotes the spatial index set of the selected brain regions, and the $\mathrm{top}-k$ operation returns the indices corresponding to the k highest values in the comprehensive score vector ${\mathbf{s}}_m^{\mathrm{pool}}$. We set $k=N/2$, meaning each pooling layer preserves 50% of the nodes [29]. Here, ${\mathbf{X}}_m^{(l+1)}$ and $A_m^{(l+1)}$ represent the pooled node feature matrix and corresponding adjacency matrix.

A soft mask mechanism is employed to smooth transitions in feature importance before node selection [30]. By weighting features with the focus scores (${\mathbf{H}}_m^{\left(l\right)}\odot {\tilde{f}}_m$) prior to $\mathrm{top}-k$ pooling, this approach (1) reduces abrupt information loss by retaining partial contributions from low-focus nodes, and (2) stabilizes training through continuous feature importance adjustment.

2.5. Hierarchical Integration Classification

Existing brain network analysis methods typically rely on the feature representation from the final layer for classification [31]. However, prior studies have shown that such a single-layer classification strategy may constrain diagnostic accuracy, as it overlooks the complementary information contained in features across different layers. Recent work further suggests that representations derived from multiple network depths offer diverse and complementary diagnostic cues for various neurological diseases [32,33].

Motivated by these findings, multi-level feature fusion has emerged as a critical strategy for improving brain disease classification. Inspired by this concept, we propose a hierarchical integration classification (HIC) framework. Specifically, we stack FAR-Conv layers and DF-Pool modules to construct a two-level architecture that captures hierarchical brain FC patterns (i.e., multi-hop interactions among ROI nodes).

The proposed network architecture comprises multiple regional processing modules connected sequentially (as illustrated in Figure 2). Each regional processing module consists of one FAR-Conv layer followed by one DF-Pool operation. The core objective of this design is to enable each module to extract brain network features at different representational levels and then systematically integrate these multi-level features for disease classification.

For the output graph $G=(V,A)$ of the l-th regional processing module, with the corresponding node feature matrix ${\mathbf{X}}^{\left(l\right)}$, a feature aggregation function is applied to obtain a fixed-size graph-level representation:

$${\mathbf{S}}^{\left(l\right)}=\mathrm{mean}\left({\mathbf{X}}^{\left(l\right)}\right)\parallel \max \left({\mathbf{X}}^{\left(l\right)}\right)$$

(13)

Here, $\mathrm{mean}\left({\mathbf{X}}^{\left(l\right)}\right)$ computes the average node features to capture global distribution patterns, while $\max \left({\mathbf{X}}^{\left(l\right)}\right)$ extracts the most salient feature values in each dimension. The concatenation operator (‖) then combines these two representations, preserving both global statistical and local discriminative information to produce a comprehensive graph-level feature embedding.

Unlike conventional methods that merely concatenate representations from different layers, we introduce an adaptive weighted integration mechanism:

$$\begin{array}{cc}\hfill {\alpha }_l& ={\displaystyle \frac{exp\left(e_l\right)}{{\sum }_{j=1}^{L}exp\left(e_j\right)}},\hfill \end{array}$$

(14)

$$\begin{array}{cc}\hfill \mathsf{\Psi }& ={\displaystyle \sum _{l=1}^L}{\alpha }_l·{\mathrm{Proj}}_l\left({\mathbf{S}}^{\left(l\right)}\right),\hfill \end{array}$$

(15)

where ${\alpha }_l$ represents the adaptive layer-wise weight computed via a softmax normalization, $e_l$ is a learnable scalar indicating the relative importance of the l-th layer, and ${\mathrm{Proj}}_l(·)$ denotes a projection function that maps features from different layers into a unified embedding space. All parameters are jointly optimized through backpropagation, enabling the network to automatically determine the contribution of each layer in the overall classification process.

Finally, the integrated feature vector $\mathsf{\Psi }$ is passed through a multilayer perceptron (MLP) followed by a softmax activation to obtain the final predicted label:

$$\widehat{y}=\mathrm{softmax}\left(\mathrm{MLP}\left(\mathsf{\Psi }\right)\right)$$

(16)

2.6. Loss Function Design

To optimize our proposed HF-BrainGNN model—particularly to guide effective learning of the hierarchical integration classifier introduced in Section 2.3—we design three complementary loss functions. These loss functions target classification accuracy, feature selection clarity, and intra-group consistency, respectively, together forming a multi-objective optimization framework.

2.6.1. Classification Loss

As the primary optimization objective, we adopt the standard binary cross-entropy loss function [34], which effectively addresses the class imbalance issue common in brain disease datasets:

$${\mathcal{L}}_{CE}=-{\displaystyle \frac{1}{M}}{\displaystyle \sum _{m=1}^M}\left[y_{m}log\left({\widehat{y}}_m\right)+(1-y_m)log\left(1-{\widehat{y}}_m\right)\right]$$

(17)

Here, $y_m\in \{0,1\}$ denotes the true class label of subject m, and ${\widehat{y}}_m$ represents the model’s predicted probability that the subject belongs to the positive class (EMCI). ${\widehat{y}}_m$ is the predicted probability output by the hierarchical integration classifier (introduced in Section 2.5).

2.6.2. Focus Separation Loss

To enhance the feature selection capability of the DF-Pool layer and guide the hierarchical integration classifier toward learning distinct and discriminative node representations, we design the focus separation (FS) loss:

$${\mathcal{L}}_{FS}=-{\displaystyle \frac{1}{M}}{\displaystyle \sum _{m=1}^M}\left({\displaystyle \frac{1}{k}}{\displaystyle \sum _{i=1}^k}{\tilde{f}}_{m,\left(i\right)}-{\displaystyle \frac{1}{N-k}}{\displaystyle \sum _{j=k+1}^N}{\tilde{f}}_{m,\left(j\right)}\right)$$

(18)

The core idea of this loss is to maximize the difference between the average focus scores of the selected (top-k) nodes and the unselected nodes. For each subject m, we first sort the normalized focus scores ${\tilde{f}}_m$ in descending order. Here, ${\tilde{f}}_{m,\left(i\right)}$ denotes the i-th highest normalized focus score. We then compute the mean score of the top-k selected nodes and that of the remaining $(N-k)$ unselected nodes, taking the negative of their difference. Minimizing ${\mathcal{L}}_{FS}$ thus explicitly enforces the pooling layer to produce a sharper contrast between informative and less informative brain regions.

2.6.3. Consistency Regularization Loss

To balance the individual variability and within-class consistency of brain disease patterns [35], we introduce the consistency regularization (CR) loss:

$${\mathcal{L}}_{CR}={\displaystyle \sum _{c=1}^C}{\displaystyle \frac{1}{|I_c{|}^2}}\sum _{m,n\in I_c}{\parallel {\tilde{\mathbf{f}}}_m-{\tilde{\mathbf{f}}}_n\parallel }_1{\mathcal{L}}_{FS}=-{\displaystyle \frac{1}{M}}{\displaystyle \sum _{m=1}^M}\left({\displaystyle \frac{1}{k}}{\displaystyle \sum _{i=1}^k}{\tilde{f}}_{m,\left(i\right)}-{\displaystyle \frac{1}{N-k}}{\displaystyle \sum _{j=k+1}^N}{\tilde{f}}_{m,\left(j\right)}\right)$$

(19)

Here, $I_c$ denotes the index set of subjects belonging to class c, and $|I_c|$ is the number of subjects in that class. This term computes the average L1-distance between the top-k node focus scores of different subjects within the same class. The L1 norm is adopted to reduce the influence of outliers and improve robustness.

By constraining subjects of the same class to maintain similar focus distributions, this regularization promotes group-level consistency. It enables the hierarchical integration classifier to form stable, class-specific attention weight patterns, thereby strengthening discriminative feature representations and improving classification performance.

2.6.4. Total Loss

We combine the three loss functions into a total loss function, where the weights of each component are adjusted via hyperparameters:

$${\mathcal{L}}_{\mathrm{total}}={\mathcal{L}}_{CE}+{\lambda }_1{\mathcal{L}}_{FS}+{\lambda }_2{\mathcal{L}}_{CR}$$

(20)

This multi-objective optimization design enables the model to pursue multiple goals simultaneously: ${\mathcal{L}}_{CE}$ ensures prediction accuracy by directly optimizing the output of the hierarchical integration classifier; ${\mathcal{L}}_{FS}$ enhances the clarity of node selection; ${\mathcal{L}}_{CR}$ balances individual specificity and group consistency, thereby improving the stability of hierarchical integration.

The hyperparameters ${\lambda }_1$ and ${\lambda }_2$ play a key balancing role: by tuning their values, we can trade off between the model’s classification performance and interpretability [36].

3. Experiments and Results

3.1. Datasets

We used public data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI; https://adni.loni.usc.edu, (accessed on 6 March 2025)), which includes resting-state functional magnetic resonance imaging (rs-fMRI) data from 104 EMCI patients and 114 cognitively normal (CN) participants. The demographic and clinical characteristics of the study participants are summarized in

Table 2. Demographic and clinical characteristics of study participants.

Characteristic	EMCI ($\mathit{n}=104$)	CN ($\mathit{n}=114$)
Age	$74.94\pm 6.11$	$74.48\pm 6.48$
Gender (M/F)	56/48	60/54
Education (years)	$15.19\pm 2.98$	$15.78\pm 3.02$
MMSE Score	$24.39\pm 2.48$	$28.12\pm 1.21$

. To quantify the cognitive status of the participants, we utilized the Mini-Mental State Examination (MMSE), a widely recognized 30-point clinical tool used to screen for cognitive impairment and dementia. MMSE scores range from 0 to 30, where higher scores indicate better cognitive function. Typically, a score of 27–30 is considered cognitively normal, while lower scores reflect varying degrees of cognitive decline. All data were acquired using a Philips Medical Systems 3.0T MRI scanner, with the following scan parameters: repetition time (TR) = 3000 ms, echo time (TE) = 30 ms, flip angle = 80°, slice thickness = 3.3 mm, and number of slices = 48.

Data preprocessing was performed using the GRETNA software package [37], with key steps including: discarding the first 10 time points, temporal slice correction, head motion correction, spatial normalization to the Montreal Neurological Institute (MNI) space (resolution: $3\times 3\times 3$ mm³), spatial smoothing (6 mm full-width at half-maximum (FWHM) Gaussian kernel), linear trend removal, and band-pass filtering (0.01–0.1 Hz). These preprocessing steps conform to the widely accepted standards for rs-fMRI data analysis [38].

3.2. Experimental Setup

All experiments in this study were implemented in the PyTorch framework (version 2.5.1), and training and testing were conducted on a workstation equipped with an NVIDIA RTX 3090Ti GPU (24 GB memory). The model architecture (Figure 2) consists of two affinity-aware regional convolution (FAR-Conv) layers and two differential focus pooling (DF-Pool) layers, forming two complete regional processing modules.

For the EMCI vs. CN classification task, we set the parameters as follows: $N=90$ (number of brain regions in the AAL template); $K^{\left(0\right)}=K^{\left(1\right)}=8$ (number of functional networks); $d^{\left(0\right)}=90$ (initial feature dimension); $d^{\left(1\right)}=32$ (output feature dimension of the first layer); $d^{\left(2\right)}=32$ (output feature dimension of the second layer); and $C=2$ (number of classes).

The pooling rate is set to 0.5, meaning each pooling layer retains 50% of the nodes; after two pooling layers, approximately 25% of the original nodes are preserved as key brain regions. The setting of $K=8$ (number of functional communities) is based on the commonly used eight functional network divisions in neuroscience research [23].

We use 5-fold cross-validation to evaluate model performance: the entire dataset is randomly partitioned into five mutually exclusive folds of approximately equal size. In each fold, four folds (80% of the data) are used as the training set (with 20% of the training data further split as the validation set for hyperparameter tuning), and the remaining one fold (20%) is used as the test set. This process is repeated five times (each fold serving as the test set once), ensuring that each subject’s data is only included in the test set once and remains independent across folds. The final performance metrics are the average of the results from the five folds.

The model is trained using the Adam optimizer, with an initial learning rate of 0.001 (halved every 30 epochs), a weight decay of 0.0005, a batch size of 32, and a maximum training epoch of 120. An early stopping strategy (training stops if validation set accuracy does not improve for 15 consecutive epochs) is applied to avoid overfitting.

For the hyperparameters of the loss functions, we tune ${\lambda }_1$ (focus separation loss coefficient) and ${\lambda }_2$ (consistency regularization loss coefficient) using the training and validation sets. Model evaluation metrics include accuracy, sensitivity, specificity, and AUC; all results are reported as the mean and standard deviation across 5-fold cross-validation.

3.3. Hyperparameter Discussion and Ablation Study

3.3.1. Hyperparameter Settings

To examine the impact of hyperparameters on model performance, we conducted a grid search to tune ${\lambda }_1$ and ${\lambda }_2$ in the loss function using the training and validation sets, with the search ranges set as ${\lambda }_1\in \{0,0.01,0.05,0.1,0.2,0.5\}$ and ${\lambda }_2\in \{0,0.01,0.05,0.1,0.2,0.5,1.0\}$. As defined in Section 2.6, ${\lambda }_1$ controls the focus separation (FS) loss, where larger values encourage more distinct focus scores, and ${\lambda }_2$ controls the consistency regularization (CR) loss, balancing individual diversity and group-level consistency. A larger ${\lambda }_2$ promotes shared group-level patterns, while a smaller one captures individual-specific variations.

To evaluate the proposed affinity-aware regional convolution (FAR-Conv), we conducted an ablation study by replacing the FAR-Conv layer with a standard graph convolution using shared kernels (Standard-Conv). The performance gap between the two reflects the effect of regional adaptivity.

We adopted a heuristic tuning procedure: first fixing ${\lambda }_1$ to 0 or 0.1 while adjusting ${\lambda }_2$, and then optimizing ${\lambda }_1$ based on the best ${\lambda }_2$. Figure 4 shows the impact of different hyperparameter combinations on validation accuracy.

3.3.2. Ablation Study

We systematically examine the effects of ${\lambda }_1$ (focus separation, FS loss) and ${\lambda }_2$ (consistency regularization, CR loss), and evaluate the contribution of FAR-Conv compared with the standard convolution version.

Figure 4 illustrates how different loss coefficients affect validation performance for both variants.

When ${\lambda }_1=0$, the accuracy remains relatively stable while ${\lambda }_2\in [0,0.5]$, but declines when ${\lambda }_2$ reaches 1.0; Figure 4a. The highest accuracy occurs at ${\lambda }_2=0.1$, indicating that moderate consistency regularization effectively alleviates overfitting, whereas excessive constraints lead to underfitting.

As ${\lambda }_1$ increases to 0.1 (Figure 4b), performance decreases when ${\lambda }_2$ exceeds 0.2. Notably, the results at ${\lambda }_2=0$ are higher than those obtained with ${\lambda }_1=0$, showing that FS loss itself introduces a mild regularization effect that stabilizes training.

After fixing ${\lambda }_2=0.1$, varying ${\lambda }_1$ between 0 and 0.5 yields the best accuracy at ${\lambda }_1=0.1$ (Figure 4c), while stronger FS constraints (${\lambda }_1=0.2$ or 0.5) slightly reduce performance, suggesting an optimal trade-off between inter-group differentiation and model flexibility.

Across all hyperparameter settings, FAR-Conv consistently surpasses Standard-Conv, as clearly shown in every subplot. This performance gap highlights FAR-Conv’s ability to assign region-specific convolution kernels, effectively capturing the functional heterogeneity of brain networks that uniform kernels overlook.

Based on the comprehensive analysis, ${\lambda }_1=0.1$ and ${\lambda }_2=0.1$ are selected as the optimal coefficient combination, achieving the highest validation accuracy. This configuration is therefore adopted for all subsequent experiments, with results evaluated on an independent test set for reliability.

3.4. Comparison with Baseline Methods

To evaluate the effectiveness of the proposed HF-BrainGNN, we compare it with both traditional machine learning models and widely used deep learning approaches on the EMCI vs. CN classification task. Traditional handcrafted methods include ridge classifier [39], support vector machine (SVM) with RBF kernel [40], and random forest with 100 estimators [41]. These models use the vectorized FC matrix as input features and are implemented via the scikit-learn library [42]. For deep learning methods, we adopt a standard deep neural network (DNN) [43]. Graph-based neural network models specifically designed for brain analysis include GCN [44], GAT [45], BrainGNN [16], MVS-GCN [46], PopulationGCN [47], and Hi-GCN [48]. It is important to note that while models like PopulationGCN [47] utilize a population-graph approach (where nodes represent subjects), our HF-BrainGNN and other baselines like Hi-GCN follow a subject-graph paradigm (where nodes represent brain regions). This allows HF-BrainGNN to focus on learning region-adaptive features and hierarchical topology within the individual brain network.

All models were evaluated on the ADNI dataset (218 subjects) under identical experimental configurations to ensure fair and comparable performance evaluation, with 5-fold cross-validation repeated five times and the final results reported as the mean ± standard deviation across all repetitions. Traditional machine learning and DNN methods adopted 8100-dimensional vectorized connectivity features as input, while all GNN-based models used brain graphs constructed from the AAL90 atlas, where edge weights were calculated via Pearson correlation and only the top 20% of the strongest connections were retained for graph topology construction. For all deep learning models, including DNN and various GNN variants, the Adam optimizer with a fixed learning rate of 0.001 was uniformly employed, and a consistent early stopping strategy was applied to mitigate overfitting. Moreover, hyperparameters for all compared models (e.g., regularization coefficients, hidden layer dimensions, top-k pooling ratios) were optimized through a unified grid search on the validation set of each cross-validation fold, ensuring that the hyperparameter tuning criteria and implementation approaches were consistent across all methods for rigorous and unbiased comparison.

Quantitative results are reported in

Table 3. Classification results of different methods on the ADNI dataset.

Method	Acc (%)	Sen (%)	Spe (%)	AUC	F1 (%)
Ridge Classifier [39]	$73.82\pm 1.07$	$73.18\pm 2.30$	$65.05\pm 1.42$	$0.626\pm 0.035$	$77.59\pm 2.89$
SVM [40]	$73.95\pm 0.68$	$80.09\pm 0.86$	$66.29\pm 2.21$	$0.675\pm 0.023$	$82.77\pm 0.82$
Random Forest [41]	$74.69\pm 1.57$	$77.48\pm 3.67$	$63.24\pm 3.07$	$0.621\pm 0.042$	$73.17\pm 2.81$
DNN [43]	$74.61\pm 4.45$	$78.62\pm 2.75$	$68.75\pm 3.85$	$0.719\pm 0.037$	$83.90\pm 2.30$
GCN [44]	$76.91\pm 2.58$	$78.72\pm 3.13$	$66.95\pm 4.40$	$0.642\pm 0.024$	$81.61\pm 1.81$
GAT [45]	$76.91\pm 1.29$	$80.04\pm 3.68$	$67.97\pm 2.15$	$0.708\pm 0.022$	$82.96\pm 0.97$
BrainGNN [16]	$77.66\pm 4.25$	$81.86\pm 3.04$	$70.88\pm 3.62$	$0.636\pm 0.052$	$82.82\pm 1.96$
MVS-GCN [46]	$69.28\pm 4.27$	$72.66\pm 3.21$	$76.26\pm 2.69$	$0.602\pm 0.049$	$71.76\pm 2.64$
PopulationGCN [47]	$79.90\pm 1.82$	$79.59\pm 1.29$	$74.15\pm 3.73$	$0.523\pm 0.015$	$86.14\pm 1.09$
Hi-GCN [48]	$82.14\pm 1.35$	$83.01\pm 0.86$	$67.71\pm 4.79$	$0.791\pm 0.030$	$86.96\pm 1.07$
HF-BrainGNN (Ours)	$86.78\pm 1.42$	$85.20\pm 1.63$	$87.30\pm 1.53$	$0.867\pm 0.014$	$89.13\pm 1.12$

, including accuracy (Acc), sensitivity (Sen), specificity (Spe), area under the curve (AUC), and F1-score. As shown in the table, traditional handcrafted methods perform poorly, with accuracies below 75% and specificity mostly under 65%. This indicates that vector-based approaches lose topological structure information, making them prone to false positives and misclassification. The DNN achieves slightly better accuracy (76.84%), benefiting from nonlinear feature learning but still suffering from the loss of spatial adjacency information inherent to vectorization.

In contrast, all GNN-based models outperform traditional and general deep learning approaches, confirming the importance of preserving brain network structure. Among these, Hi-GCN achieves 82.14% accuracy—the best among existing baselines—while other GNN variants (GAT 78.52%, BrainGNN 79.81%, PopulationGCN 80.23%) follow closely behind.

From Table 3, the proposed HF-BrainGNN exhibits the highest overall performance: Accuracy = $86.78\%$, Sensitivity = $85.20\%$, Specificity = $87.30\%$, and AUC = $0.867$. The sensitivity improvement is particularly notable, demonstrating superior capability to correctly identify EMCI patients. High sensitivity is essential in clinical diagnosis to minimize false negatives, ensuring that patients with early mild cognitive impairment are detected reliably.

Compared with Hi-GCN, HF-BrainGNN improves accuracy by 4.64% and specificity by 15.82%, while achieving a 21.01% higher specificity than SVM (the best traditional method). Furthermore, a paired t-test evaluated across the cross-validation folds confirms that the overall performance improvement of HF-BrainGNN over the best-performing baseline (Hi-GCN) is statistically significant ($p<0.05$). These improvements are mainly attributed to: 1. The region-adaptive convolution (FAR-Conv), capturing distinct connectivity patterns across different brain areas; 2. The disease-oriented differential focus pooling (DF-Pool), emphasizing discriminative regions related to EMCI; 3. The hierarchical feature integration mechanism, which leverages complementary multi-level information.

Overall, HF-BrainGNN consistently outperforms both conventional and existing GNN-based baselines, achieving higher diagnostic sensitivity and reducing false positives—an essential property for reliable clinical applications.

3.5. Interpretability of HF-BrainGNN

3.5.1. Biomarker Detection

To verify the neurobiological significance of the key brain regions identified by HF-BrainGNN, we analyzed the salient ROIs automatically selected by the differential focus pooling (DF-Pool) layer and compared them with previously reported EMCI biomarkers. The average node focus scores of twenty-three salient brain regions were ranked across all EMCI subjects in descending order. Specifically, these scores represent the mean importance values derived from the test subjects across all 5-fold cross-validation runs. And the top ten regions were then selected as core biomarkers for in-depth analysis. Their differential focus scores ranged from 0.52 to 0.88 (

Table 4. Ten core brain regions identified by HF-BrainGNN.

Rank	AAL Abbreviation	AAL-ID	Differential Score	Functional Network
1	HIP.L	37	0.88	Hippocampal Memory System
2	HIP.R	38	0.84	Hippocampal Memory System
3	PCUN.L	67	0.79	Default Mode Network
4	PCG.R	36	0.75	Default Mode Network
5	ANG.L	65	0.71	Semantic Network
6	MTG.L	85	0.67	Semantic Network
7	PHG.R	40	0.63	Hippocampal Memory System
8	SFGmed.L	23	0.59	Executive Control Network
9	ACG.R	32	0.55	Salience Network
10	IPL.L	61	0.52	Attention Network

), and the variation of regional importance is shown in Figure 5b.

The ten most discriminative brain regions identified by HF-BrainGNN (Figure 5a) include the left and right hippocampus (HIP.L, HIP.R), left precuneus (PCUN.L), right posterior cingulate gyrus (PCG.R), left angular gyrus (ANG.L), left middle temporal gyrus (MTG.L), right parahippocampal gyrus (PHG.R), left medial superior frontal gyrus (SFGmed.L), right anterior cingulate gyrus (ACG.R), and left inferior parietal lobule (IPL.L). These regions show a function-oriented distribution pattern, with the hippocampus exhibiting the strongest memory-related characteristics.

To examine the stability of DF-Pool across folds, we extracted the top-five nodes per fold in a five-fold cross-validation scheme and constructed a binary “ROI × Fold” selection matrix ($1=\mathrm{selected}$, $0=\mathrm{not}\mathrm{selected}$). As shown in Figure 6, regions such as the hippocampus (37 HIP.L, 38 HIP.R) were repeatedly chosen across folds, demonstrating high reproducibility and stable identification of EMCI-related biomarkers. It is worth noting that while HIP.L achieved the highest global differential score (Table 4), local data distributions within individual cross-validation folds cause minor fluctuations in the fold-specific top-5 rankings. This behavior is expected and accurately reflects the natural intersubject variability in functional connectivity.

Most of the top regions are functionally linked to early EMCI pathology. In the hippocampal memory system, the bilateral hippocampus and right parahippocampal gyrus serve as core components for memory encoding and consolidation; their functional disconnection from the posterior cingulate gyrus, as reported by Sorg et al. [49], correlates with memory decline. HF-BrainGNN accurately captured these regions, validating its ability to identify memory-related impairments. Within the Default Mode Network (DMN), the precuneus and posterior cingulate gyrus act as integrative hubs for self-referential processing. Greicius et al. [50] demonstrated that DMN abnormalities distinguish healthy elders from AD patients; our results reveal that such disruptions can already be detected at the EMCI stage. In the semantic network, the left angular gyrus and left middle temporal gyrus are involved in semantic memory storage and retrieval, and their reduced connectivity strength may explain early semantic deficits in EMCI patients.

Overall, these findings show strong agreement with previous EMCI studies, confirming the neurobiological relevance of the identified regions. Reduced hippocampal connectivity [49], DMN hub abnormalities [51], and weakened angular gyrus connectivity [23] align closely with our model outcomes. Additionally, the right parahippocampal gyrus, left medial superior frontal gyrus, right anterior cingulate gyrus, and left inferior parietal lobule identified by HF-BrainGNN have all been reported as important biomarkers in prior EMCI studies [52,53,54,55]. This high consistency validates the scientific rigor and physiological plausibility of our approach, providing reliable imaging biomarkers for early EMCI diagnosis and prognosis evaluation.

Furthermore, to investigate the clinical validity and interpretability of the learned focus scores, we performed a Pearson correlation analysis between the subject-level focus scores (pooled across the top-five regions) and the clinically established MMSE scores. As illustrated in Figure 7, a strong, statistically significant negative correlation was observed ($r=-0.63$, $p<0.001$). This confirms that higher focus scores assigned by HF-BrainGNN effectively correspond to lower cognitive function (i.e., lower MMSE scores). Ultimately, this robust alignment between our model-derived metrics and standard clinical cognitive assessments underscores the clinical relevance and interpretability of the proposed framework.

3.5.2. Functional Connectivity Analysis of Core Brain Regions

To further validate the pathological significance of the core brain regions identified by HF-BrainGNN, we examined the differential patterns of twelve key functional connections constructed from the ten core biomarker regions between the CN and EMCI groups. As illustrated in Figure 8, these twelve connections can be categorized as follows: intra-connections within the hippocampal memory system (HIP.L–HIP.R, HIP.L–PHG.R); hippocampus–Default Mode Network (DMN) links (HIP.L–PCUN.L, HIP.R–PCG.R); intra-connections of the DMN (PCUN.L–PCG.R, PCUN.L–ANG.L, PCG.R–SFGmed.L, ACG.R–PCUN.L); semantic cognition network connections (ANG.L–MTG.L, MTG.L–PHG.R); and executive control links (SFGmed.L–ACG.R, IPL.L–ANG.L).

Statistical analyses revealed that the FC strength of all twelve pairs exhibited varying degrees of reduction in the EMCI group. The significance of these differences was evaluated using independent-sample t-tests with FDR correction for multiple comparisons (q < 0.05). Among these, connections in the hippocampal memory system showed the most pronounced impairment: the bilateral hippocampal link (HIP.L–HIP.R) was significantly weakened in EMCI subjects ($\mathrm{CN}:0.20\pm 0.15\mathrm{vs}.\mathrm{EMCI}:0.12\pm 0.18$, $p<0.001$), and the right hippocampus–posterior cingulate gyrus connection (HIP.R–PCG.R) also declined markedly ($\mathrm{CN}:0.25\pm 0.12\mathrm{vs}.\mathrm{EMCI}:0.15\pm 0.14$, $p<0.001$). Intra-DMN connections were similarly reduced—for example, the precuneus–posterior cingulate gyrus link (PCUN.L–PCG.R) decreased significantly in EMCI ($\mathrm{CN}:0.35\pm 0.10\mathrm{vs}.\mathrm{EMCI}:0.25\pm 0.12$, $p<0.01$).

In contrast, semantic and executive control connections appeared relatively preserved. Several pairs, such as HIP.L–PCUN.L, PCUN.L–ANG.L, MTG.L–PHG.R, and SFGmed.L–ACG.R, did not show significant group differences, suggesting partial resistance to early-stage impairment or the engagement of compensatory mechanisms in EMCI.

This distinct pattern of connectivity reduction highlights the vulnerability of hippocampal and DMN networks and supports the clinical relevance of HF-BrainGNN-identified regions. Collectively, these results provide compelling network-level evidence for understanding the neuropathological mechanisms underlying early mild cognitive impairment.

To further quantify the magnitude of these differences, we calculated Cohen’s d as a measure of effect size for the twelve identified connections, with values ranging from 0.32 to 0.91. Specifically, the connection between the bilateral hippocampus (HIP.L-HIP.R) showed a medium effect size ($d=0.49$), while the precuneus-posterior cingulate gyrus connection (PCUN.L-PCG.R) exhibited a large effect size ($d=0.91$). These results indicate that the identified FC reductions are not only statistically significant but also represent substantial biological alterations in the EMCI brain.

4. Discussion

The proposed hierarchical focused brain graph neural network (HF-BrainGNN) demonstrates clear advantages in classifying early mild cognitive impairment (EMCI) versus the control normal (CN), achieving an accuracy of 86.78%, substantially surpassing existing graph-based and conventional machine-learning methods. These improvements arise from the synergistic effects of its three key components.

First, the FAR-Conv layer introduces region-specific kernels by integrating functional similarity and anatomical proximity, effectively addressing the limitation that traditional GNNs neglect regional functional heterogeneity. Ablation experiments confirm that FAR-Conv consistently outperforms standard graph convolution across parameter configurations, validating the benefit of region-adaptive modeling. This mechanism accords with the neuroscientific principle of functional specialization, offering a biologically plausible representation framework for brain connectivity analysis.

Second, the DF-Pool layer performs disease-oriented feature selection by comparing inter-group focus scores to automatically identify EMCI-relevant brain regions. The ten salient biomarkers discovered, covering the hippocampal memory system, the default mode network, and the semantic network, align closely with previous EMCI studies and exhibit negative correlations with MMSE scores ($r=-0.63,p<0.001$). This result furnishes an interpretable neurobiological basis for model predictions.

Beyond biomarker identification, the hierarchical integration classifier (HIC) adaptively fuses multi-level representations through learnable weights, capturing multi-scale functional interactions within the brain network. Together with a multi-objective loss combining classification accuracy, focal separation, and consistency regularization, HF-BrainGNN achieves an optimal balance between performance and interpretability, strengthening its potential for clinical translation.

Furthermore, the results of this study are dependent on the AAL90 parcellation scheme used to define the graph nodes. While AAL90 is a widely accepted anatomical reference, it may introduce parcellation bias, as anatomical boundaries do not always align perfectly with functionally specialized areas. Future research should evaluate the robustness of HF-BrainGNN across diverse brain templates, such as functional or higher-resolution multi-scale atlases, to mitigate the impact of atlas selection on diagnostic performance.

Nevertheless, several directions remain open for future investigation. While HF-BrainGNN demonstrated superior performance on the ADNI dataset, its generalizability across different imaging protocols and diverse populations has not yet been fully established. Expanding the dataset to multi-center cohorts and cross-scanner acquisitions will help evaluate model generalization and ensure the robustness of the identified biomarkers. Furthermore, extending HF-BrainGNN to other neuropsychiatric conditions, such as Parkinson’s disease or depression, may broaden its utility for early diagnostic and prognostic imaging biomarker discovery.

5. Conclusions

This work presents HF-BrainGNN, a hierarchical graph neural network for EMCI diagnosis based on rs-fMRI brain networks, which incorporates novel designs, including FAR-Conv, DF-Pool, and HIC, to address the functional heterogeneity, group-level feature selection, and multi-level feature utilization problems in existing methods. Experiments on the ADNI dataset demonstrate that HF-BrainGNN achieves outstanding performance, with a classification accuracy of 86.78%, sensitivity of 85.2%, and specificity of 87.3%. The identified biomarkers are consistent with existing neuroscience findings, ensuring both algorithmic performance and interpretability. This study provides a new technical framework for neuroimaging-based disease diagnosis and an interpretable AI tool for early EMCI diagnosis, with important implications for AD early intervention. Future work will focus on multi-center dataset validation, MCI progression analysis, and cross-disease application of the HF-BrainGNN framework to advance brain imaging precision medicine.