SpeechHGT: A Multimodal Hypergraph Transformer for Speech-Based Early Alzheimer’s Disease Detection

1 Introduction

Early diagnosis of Alzheimer's disease (AD) is crucial for timely intervention and improved patient outcomes [Alberdi et al., 2016; Shehzad et al., 2025]. AD is a neurodegenerative disorder characterized by memory loss, cognitive decline, and behavioral changes [Marvi et al., 2024; Zhang et al., 2024]. While neuroimaging techniques like MRI and PET can detect brain alterations, their utility is limited by high costs, restricted accessibility, and radiation exposure from repeated PET scans [Ahmed et al., 2019; Yu et al., 2024; Yang et al., 2022]. Consequently, there is a growing need for cost-effective, non-invasive diagnostic methods [Petti et al., 2020; Ding et al., 2024]. Speech analysis shows promise for early AD detection, leveraging both linguistic (e.g., word choice, syntactic complexity) and acoustic (e.g., speech rate, pitch) features [Pulido et al., 2020; Pacheco-Lorenzo et al., 2024]. However, effectively integrating these diverse speech features to capture the intricate patterns of cognitive decline necessitates further research and advanced modeling approaches.

∗Corresponding author.

Previous methods in speech-based AD detection can be categorized into unimodal and multimodal approaches, each employing distinct computational methodologies [Latif et al., 2021; Shehzad et al., 2024]. Acoustic methods utilize prosodic features, such as pitch, formant frequencies, and temporal variations, to identify vocal anomalies linked to ADrelated neurodegeneration [Luz et al., 2024; Zhang et al., 2021]. Linguistic approaches focus on lexical, syntactic, and semantic features, examining word frequency, sentence structure, and narrative coherence to detect cognitive impairments. However, unimodal methods often neglect cross-modal interactions, leading to incomplete assessments and reduced diagnostic accuracy. Therefore, multimodal architectures integrating information from multiple sources are adopted [Vrindha et al., 2023; Venugopalan et al., 2021]. These systems integrate acoustic and linguistic representations through hierarchical fusion strategies [Turrisi et al., 2024]. Recent advances employ graph transformers to model complex intermodal relationships in multimodal data, enhancing diagnostic performance [Ektefaie et al., 2023; Peng et al., 2024]. By leveraging attention mechanisms, these models effectively capture intricate feature dependencies, demonstrating significant potential for improving AD classification from spontaneous speech [Bessadok et al., 2022].

Despite advancements, current multimodal speech analysis techniques often miss crucial, complex interactions between linguistic and acoustic features-interactions vital for early Alzheimer's disease (AD) detection [Ying et al., 2023; Priyadarshinee et al., 2023]. These interactions, like the interplay of pitch, tempo, and prosody, reflect cognitive impairments in AD. Traditional methods, such as linear aggregation, typically treat these features independently, assuming simple additive relationships [Ilias and Askounis, 2022a]. This overlooks the inherent non-linear dependencies in speech data [P´erez-Toro et al., 2021], potentially leading to delayed or inaccurate diagnoses. We hypothesize that incorporating these higher-order interactions will significantly improve diagnostic accuracy, aligning with neurolinguistic theories of cognitive decline [Dell, 1986]. This study aims to validate this hypothesis and enhance non-invasive diagnostic tools for early AD detection.

To address these challenges, we propose SpeechHGT 1 , a multimodal hypergraph transformer, to model higher-order interactions between linguistic and acoustic speech features for improved AD detection. SpeechHGT extracts discriminative features from preprocessed audio, representing ADrelated speech characteristics. We construct a multimodal hypergraph where nodes denote individual features, and hyperedges capture grouped interactions. This hypergraph integrates both simple edges for pairwise relations and hyperedges for higher-order dependencies. To process this hypergraph-structured data, we design a novel hyperedge attention-based transformer model, which captures both pairwise and higher-order interactions. Transformed node features are aggregated for binary classification, distinguishing AD from speech samples. Experimental results demonstrate that SpeechHGT outperforms baseline models in accuracy and F1-score, offering an effective approach for early AD detection and improved diagnostic reliability.

Our contributions are as follows.

1. We propose SpeechHGT, a novel multimodal hypergraph transformer that captures higher-order interactions between linguistic and acoustic speech features, overcoming the limitations of existing fusion-based approaches for AD detection in capturing complex dependencies.
2. We design a dual-layer hypergraph attention mechanism that effectively models both pairwise and higher-order dependencies, which can improve the integration of multimodal speech features for robust classification.
3. Extensive experiments on multiple real-world datasets show that SpeechHGT outperforms state-of-the-art methods in speech-based AD classification. It achieves higher accuracy, and F1-score on all benchmark datasets, demonstrating its effectiveness in improving early AD diagnosis from spontaneous speech.

2 RelatedWork

2.1 Speech Analysis for Brain Disease Diagnosis

The diagnosis of speech-based neurodegenerative diseases traditionally relies on acoustic or linguistic representations [Luz et al., 2024]. Acoustic methods analyze prosodic and voice quality characteristics, including pitch contours, speaking rate, and jitter, to identify early markers of AD. [Luz et al., 2020] proposes a standardized acoustic preprocessing framework, demonstrating that prosodic indices alone reveal measurable cognitive impairment. Linguistic methods, in contrast, focus on transcribed speech, examining lexical diversity, syntactic complexity, and semantic coherence. [Searle et al., 2020] shows that advanced language embeddings, such as DistilBERT [Sanh, 2019], can enhance detection accuracy in machine learning models using textual transcripts. Despite their utility, unimodal approaches fail to integrate prosodic and linguistic features, limiting the exploration of holistic speech characteristics that are essential for comprehensive diagnostic assessments.

1 The source codes are available at: https://github.com/ Ahsan-Shehzad/SpeechHGT.

Current multimodal speech analysis techniques aim to enhance diagnostic accuracy by integrating acoustic and linguistic features using early or late fusion strategies [Ilias and Askounis, 2022b]. [Martinc and Pollak, 2020] shows that optimized combinations of text and audio outperform unimodal approaches in AD detection. Multimodal deep learning models, such as BiLSTM or Transformer architectures, improve feature integration by combining acoustic waveforms with textual transcripts. [Rohanian et al., 2021] demonstrates that gating mechanisms in sequence models align prosodic and lexical-semantic features to enhance predictions. [Zhu et al., 2021] refines semantic embeddings using non-semantic features, like pause duration, via Wav2vec. However, existing methods often fail to model complex intermodal relationships and higher-order dependencies, neglecting critical biomarkers such as semantic confusion and speech disfluencies in AD.

2.2 Graph Transformers

Graph transformers integrate the representational power of graph neural networks (GNNs) with attention-based Transformer mechanisms to model relational data [Liu et al., 2021]. These architectures propagate node-level information across structured connections while using attention coefficients to weight node or edge importance. Recent advancements in protein folding and language modeling demonstrate their ability to address complex relational data domains [Ying et al., 2021]. In clinical research, graph-based methods model disease progression, predict pathological links, and identify biomarkers [Luo et al., 2024]. Their strength lies in capturing long-range dependencies while preserving structural information. However, applying Graph Transformers to multimodal data presents challenges, such as heterogeneous feature spaces and sparse cross-modal relationships [Li et al., 2024].

3 Design of SpeechHGT

3.1 ProblemFormulation

This study proposes SpeechHGT, a multimodal hypergraph transformer, for early AD detection using DementiaBank speech data. Speech features include linguistic (FL) and acoustic (FA), forming a combined set F = FL ∪ FA. A hypergraph H = (V, E) models higher-order feature interactions, where nodes (V) represent features and hyperedges (E) encode relationships. SpeechHGT learns node (hv) and hyperedge (he) embeddings via hypergraph attention, outputting a binary classification (y ∈ {0, 1}) for AD presence. The model optimizes accuracy by minimizing the loss function L. This approach enhances AD detection and provides insights into cognitive decline. Figure 1 illustrates the framework.

Type-Token Ratio (TTR), Part-of-Speech (POS) tag distributions, Brunet's Index, and Honore's Statistic. Syntactic features encompass sentence complexity, grammatical correctness, parsing tree depth, and clause-to-sentence ratios. Semantic features, derived from contextual embeddings (e.g., BERT), evaluate coherence, semantic similarity, and named entity detection. We use Latent Dirichlet Allocation (LDA) for topic modeling. Discourse analysis involves pausing patterns, pronoun usage, narrative coherence, and topic maintenance. All features are consolidated into a structured vector Li for each audio sample.

Acoustic Feature

We analyze audio signals to capture prosodic, articulatory, and spectral properties of speech. Using LibROSA and OpenSMILE 10 , we extract phonation features, including jitter, shimmer, Harmonics-to-Noise Ratio (HNR), and Cepstral Peak Prominence (CPP), which reflect vocal stability and clarity. Temporal features, such as speaking rate, silent and filled pause durations, and turn-taking timing, characterize fluency and rhythm. Spectral features, including formant frequencies (F1, F2), Mel-Frequency Cepstral Coefficients (MFCCs), spectral slope, flux, centroid, and bandwidth, describe spectral energy distribution and dynamics. Energybased features, such as Zero-Crossing Rate (ZCR), intensity contours, sub-band energy distribution, and loudness profiles, quantify energy variations. All acoustic features are encapsulated into a feature vector Ai, providing a comprehensive representation of the audio signal.

3.6 Multimodal Hypergraph Construction

Defining Nodes

Each node in the hypergraph represents a unique linguistic or acoustic feature extracted from the DementiaBank dataset. We define the set of nodes as V = {v1, v2, . . . , vn}, where each node vi corresponds to a specific feature Fi. Each feature is assigned a unique identifier, such as F1 for Vocabulary Richness or F2 for Pitch Variability. The attribute vector ai for each node vi includes the feature type (linguistic or acoustic) and its statistical properties, specifically the mean (µi) and variance (σ2 i ). Accordingly, the feature vector for node vi is expressed as:

10 https://audeering.com/opensmile/

3.2 Audio Preprocessing

We preprocess raw audio to standardize data quality for reliable feature extraction and classification. This includes noise reduction, volume normalization, segmentation, augmentation, and transcription. All recordings are converted to WAV, resampled to 16 kHz with 16-bit depth, and set to mono using SoX 2 . NoisePy 3 applies spectral subtraction-based noise reduction using STFT-estimated noise spectrum N(f). RMS normalization ensures uniform amplitude. Voice activity detection (VAD) via py-webrtcvad 4 segments recordings. These steps improve analytical robustness.

3.3 Data Augmentation

We apply data augmentation to enhance diversity and model robustness. Speed perturbation adjusts playback speed (α ∈ {0.9, 1.0, 1.1}) while preserving pitch. Pitch shifting modifies frequency components via the semitone factor β (β ∈ {-2,-1, 1, 2}). Gaussian noise n(t) ∼ N(0, σ2 ) is added, maintaining a 20 dB SNR. Augmentations are implemented using librosa 5 .

3.4 Transcription

We use Whisper 6 , an ASR model, to transcribe audio x(t) into text T [Liu et al., 2024]. It captures filler words, disfluencies, and unintelligible segments: T = fASR(x(t);Θ), where fASR is the model and Θ its parameters. This transcription supports linguistic analysis and feature extraction.

3.5 Feature Extraction

Linguistic Feature

We analyze speech transcripts for cognitive and communicative disruptions linked to Alzheimer's Disease (AD). Utilizing advanced Natural Language Processing (NLP) tools like spaCy 7 , NLTK 8 , and Transformers 9 , we extract a wide range of linguistic features. Lexical analysis includes word count,

2 http://sox.sourceforge.net/

3 https://github.com/noisepy/NoisePy

4 https://github.com/wiseman/py-webrtcvad

5 https://librosa.org

6 https://github.com/openai/whisper

7 https://spacy.io/

8 https://www.nltk.org/

9 https://huggingface.co/docs/transformers

Identifying Hyperedges

We identify hyperedges to capture higher-order interactions influencing AD detection through statistical correlation analysis and clustering. First, we compute Pearson correlation coefficients (ρij) for all feature pairs (Fi, Fj):

Next, we apply the spectral clustering algorithm to the correlation matrix to cluster features with high inter-correlations. Let C = {C1,C2, . . . ,Cm} denote the resulting set of clusters. Each cluster Ck corresponds to a hyperedge ek defined as:

This method ensures that hyperedges represent synergistic feature groups with collective relevance to AD detection.

Incorporating Pairwise Edges

We incorporate pairwise edges to capture direct interactions between individual feature pairs, complementing hyperedges. For a given feature pair (Fi, Fj), we compute the mutual information I(Fi; Fj) to quantify feature dependency:

A pairwise edge is established between nodes vi and vj if I(Fi; Fj) exceeds a predefined threshold θ. These pairwise edges capture direct feature dependencies absent in higherorder groupings, enhancing the hypergraph's structural complexity and representational depth.

Representing the Hypergraph

We represent the hypergraph using an incidence matrix H ∈ {0, 1}n×m, where n is the number of nodes and m is the number of hyperedges. Each matrix element Hi,k is defined as:

We encode nodes and hyperedges with feature vectors to facilitate learning within the hypergraph framework. Each node vi is represented by a normalized feature vector xi, defined as:

Hyperedge features are aggregated from constituent node vectors. The aggregated feature vector yk for hyperedge ek is computed as:

This encoding captures collective node information, enabling the multimodal hypergraph transformer (SpeechHGT) to learn complex feature interactions. The hypergraph construction module outputs the incidence matrix H, node features {xi}ni=1 , and hyperedge features {yk}mk=1 .

3.7 Multimodal Hypergraph Transformer

The Hypergraph Transformer constitutes the core of our SpeechHGT architecture, designed to leverage hypergraphstructured data for capturing pairwise and higher-order feature interactions in spontaneous speech.

Hypergraph Attention Mechanism

We design the hypergraph attention mechanism to extend standard self-attention for modeling higher-order interactions. Our approach employs dual attention layers: NodeLevel Attention and Hyperedge-Level Attention. The NodeLevel Attention captures pairwise node interactions, formulated as:

where Q, K, and V denote the query, key, and value matrices, and dk represents the key vector dimensionality. To complement this, we implement Hyperedge-Level Attention, which aggregates features across hyperedges to capture higher-order dependencies:

Here, Qe, Ke, and Ve correspond to hyperedge-specific matrices, with de as the dimensionality of hyperedge key vectors. This dual-layer design allows us to capture both direct and collective feature interactions effectively.

Aggregate Hyperedge Embeddings

We represent the collective influence of hyperedges on connected nodes using an iterative embedding update mechanism. Each hyperedge e ∈ E starts with an initial embedding h(0) e , which we iteratively refine based on connected node features. The embedding update rule is defined as:

Here, We and be are trainable parameters, σ represents the activation function, and h(l) v denotes the node embedding at layer l. This mechanism ensures that hyperedge embeddings effectively capture aggregated information from their associated nodes.

Structural Positional Encodings

We incorporate structural positional encodings to capture nodes' structural relationships and positional contexts within the hypergraph. These encodings preserve structural integrity and enable the learning of positional dependencies. Each node's degree, defined by the number of hyperedges it participates in, is encoded as: pdegree v = Linear (log(1+ deg(v))) , where deg(v) represents the degree of node v, and Linear denotes a linear transformation. We also encode centrality measures, including betweenness (β(v)) and closeness (γ(v)) centrality, as: pcentrality v = Linear (β(v), γ(v)) . The final positional encoding combines degree and centrality encodings:

We integrate these encodings into node representations, enhancing the Transformer's ability to leverage structural information. This approach improves the model's ability to capture nuanced feature interactions essential for Alzheimer's disease detection.

Classification

The Classification module utilizes transformed node features from the Hypergraph Transformer to perform binary classification of AD in speech samples. It converts node embeddings, which capture linguistic and acoustic interactions, into a unified graph-level representation z through an attentionbased readout function:

where αv is the attention weight for node v, calculated as:

The attention mechanism ensures that significant nodes have a larger influence on z. The aggregated representation z is then processed through a fully connected layer followed by a sigmoid function to produce the probability ˆ y of AD:

The model is optimized using binary cross-entropy loss. The pseudocode of SpeechHGT is given in Algorithm 1.

Algorithm 1 SpeechHGT Algorithm

Input: Raw speech audio X

Output: AD prediction ˆ y

• 1: procedure SPEECHHGT(X)
• 2: XP ← Preprocess(X)
• 3: XA ←Augment(XP )
• 4: T ←Transcribe(XA)
• 5: L←ExtractLinguisticFeatures(T)
• 6: A←ExtractAcousticFeatures(XA)
• 7: V ←L∪A
• 8: E ←HypergraphConstruction(V)
• 9: H0 ← InitializeEmbeddings(V, E)
• 10: HS ←ApplyStructuralEncodings(H0, E)
• 11: Initialize Transformer weights θ
• 12: E ←number of training epochs
• 13: for e ← 1 to E do
• 14: HN ←NodeLevelAttention(HS, E)

15: HE ←HyperedgeLevelAttention(HN, E)

• 16: end for
• 17: Z←AggregateEmbeddings(HE)
• 18: ˆ y ← σ(Wn . . . σ(W1Z + b1) · · · + bn)
• 19: Compute loss Lcls
• 20: Update θ using backpropagation
• 21: return ˆ y
• 22: end procedure

4 Experiments

4.1 Experimental Settings

Datasets

We evaluate SpeechHGT using three benchmark datasets from the DementiaBank repository: ADReSS, ADReSSo, and ADReSS-M. These datasets, derived from the Cookie Theft Picture Description Task, are designed for AD detection based on spontaneous speech data. Table 1 summarizes their key characteristics, with details provided below.

• The ADReSS dataset [Luz et al., 2020] includes 156 samples (78 AD, 78 cognitively normal [CN]) with balanced gender representation (44.9% male, 55.1% female) and a mean age of 66.8 years. A 70/30 train-test split is employed, but its small size limits robust training and increases overfitting risk.
• The ADReSSo dataset [Luz et al., 2021] contains 237 samples (122 AD, 115 CN), with a higher female representation (34.9% male, 65.1% female) and greater age variability. The mean ages are 69.38 years (AD) and 66.06 years (CN). Its moderate size and demographic diversity enable model evaluation under variable conditions.
• The ADReSS-M dataset [Luz et al., 2024], the largest, comprises 271 samples (132 AD, 139 CN), with a gender distribution of 33.6% male and 66.4% female. The dataset uses an 80/20 train-test split, offering stability for model training while presenting demographic imbalance challenges.

4.2 Baselines

Challenge Baselines

These correspond to the methodologies established in the ADReSS, ADReSSo, and ADReSS-M challenges, which serve as standardized benchmarks for AD classification and cognitive score prediction.

Conventional Machine Learning Models

We implement Random Forest (RF), Support Vector Machines (SVM), and AdaBoost, focusing on linguistic features extracted from speech data. These models benchmark classical techniques against deep learning methods.

Unimodal Speech Models

These models utilize either linguistic or acoustic features for Alzheimer's detection. For instance, [Searle et al., 2020] employ TF-IDF and DistilBERT embeddings, while [P´erezToro et al., 2021] use X-vectors, prosody, and emotional embeddings. We replicate their feature extraction pipelines and training procedures.

Multimodal Speech Models

These models integrate linguistic and acoustic features to address AD-related speech complexities. The baselines include [Martinc and Pollak, 2020], [Rohanian et al., 2021], [Zhu et al., 2021], [Chen et al., 2023], [Tamm et al., 2023], [Lin and Washington, 2024]. We adhere to the architectures and parameter configurations described in their experiments.

4.3 Implementation Details

We implement the SpeechHGT framework using the PyTorch Geometric library to model multimodal hypergraphs efficiently. The architecture includes two key modules: multimodal hypergraph construction and the multimodal hypergraph transformer. We evaluate the framework on ADReSS, ADReSSo, and ADReSS-M datasets, training and testing separately. The training uses the Adam optimizer (learning rate: 0.001, batch size: 32) with early stopping based on validation loss (patience: 10 epochs). Hyperparameters, including attention heads, hidden dimensions, and dropout rates, are optimized via grid search. All experiments utilize an NVIDIA RTX 4090 GPU, Intel i9 13th Gen CPU, and 64GB RAM, ensuring scalability and computational efficiency. This implementation achieves robust higher-order interaction modeling, validating our framework's effectiveness.

4.4 Performance of SpeechHGT

We evaluate SpeechHGT on ADReSS, ADReSSo, and ADReSS-M datasets using standard binary classification metrics. Results (Table 2) confirm high precision and reliability in AD detection. On ADReSS, the model achieves an accuracy of 86.32%, precision of 86.14%, recall of 85.28%, and F1-score of 86.69%. For ADReSSo, it attains an accuracy of 88.18%, with precision of 89.27% and recall of 88.54%, leveraging hypergraph attention for linguistic-acoustic dependencies. Despite greater heterogeneity in ADReSS-M, it maintains an accuracy of 82.82%, demonstrating robustness while identifying areas for improved adaptation to outlier speech patterns.

4.5 Comparison with Baseline Methods

The proposed SpeechHGT model demonstrates consistent superiority across the ADReSS, ADReSSo, and ADReSSM datasets compared to four baseline categories: challenge baselines, conventional machine learning models (e.g., RF, SVM, AdaBoost), unimodal speech models, and stateof-the-art multimodal approaches (Table 2). On average, SpeechHGT achieves 85.77% accuracy, 86.33% precision, 85.74% recall, and 85.97% F1-score, outperforming the bestperforming model in each baseline group. Compared to challenge baselines (62.50% accuracy), SpeechHGT improves performance by 23.27% on average. Against conventional ML models (best: 65.39% accuracy), it achieves gains of 20.38-24.98%. Unimodal models like [P´erez-Toro et al., 2021] are outperformed accuracy by 10.18% and F1-score by 7.86%. Finally, SpeechHGT surpasses state-of-the-art multimodal baselines, improving F1-score by 4.26% on average. While SpeechHGT slightly underperforms [Lin and Washington, 2024] in accuracy on ADReSS-M (-0.62%), it outperforms them in F1-score (+1.67%) and demonstrates superior consistency across all datasets, highlighting its robustness and generalizability.

Table 2: Performance comparison with different baselines (%).

4.6 Ablation Study

We conducted an ablation study to evaluate SpeechHGT framework components by disabling specific elements. First, we removed hyperedge-level attention, limiting interactions to pairwise features. Next, hyperedges were eliminated, reducing the graph to pairwise connections. We also assessed modality-specific contributions by excluding linguistic and acoustic features individually. The impact of structural positional encodings (e.g., node degrees, centrality) was also evaluated. Comparisons with a multimodal baseline highlighted their significance. As shown in Table 3, ADReSS accuracy dropped from 86.32% to 78.55% without hyperedge-level attention, and further to 76.82% without hyperedges. Removing linguistic and acoustic features resulted in 81.14% and 80.28% accuracies, respectively, while eliminating structural positional encoding led to an 82.87% accuracy.

4.7 Analysis of Higher-Order Interactions

We employ the SpeechHGT model to systematically identify and quantify higher-order interactions indicative of Alzheimer's Disease (AD)-related cognitive decline. This model utilizes a hypergraph attention mechanism, which effectively prioritizes clinically significant speech features while concurrently minimizing extraneous noise. Key results, illustrated in Figure 4, underscore the pivotal role of these interactions, such as semantic coherence and narrative structuring, in AD detection. Identified disruptions in logical flow and prosodic control align with neurolinguistic theories linking conceptual organization and motorspeech processes to cognitive decline [Rumelhart et al., 1986; Dell, 1986].

4.8 Parameter Analysis

The selection of hyperparameters significantly impacts SpeechHGT's ability to model higher-order interactions in spontaneous speech. We analyze the number of clusters (k) for hyperedge construction and transformer depth (l) for hierarchical representation learning. A grid search across k ∈ [2, 10] and l ∈ [2, 10] on ADReSS, ADReSSo, and ADReSS-M datasets identifies dataset-specific optima. Fixed hyperparameters (learning rate: 10-4 , batch size: 16, attention heads: 8, dropout: 0.3) isolate the effects of k and l. As illustrated in Figure 2 and Figure 3. Optimal cluster counts (k∗ ) are k = 3 for ADReSS and ADReSS-M and k = 4 for ADReSSo, with higher k reducing accuracy due to hyperedge fragmentation. Transformer depth (l∗ ) varies, favoring l = 4 for ADReSS, l = 2 for ADReSSo, and l = 6 for ADReSS-M, reflecting dataset-size dependencies. These results confirm SpeechHGT's sensitivity to parameter tuning across heterogeneous datasets.

5 Conclusion

This study presents SpeechHGT, a novel multimodal hypergraph transformer designed to address the limitations of existing fusion-based models in AD detection through spontaneous speech analysis. By introducing a hypergraph-based approach to represent and learn higher-order interactions between linguistic and acoustic features, SpeechHGT achieved significant improvements in diagnostic accuracy, F1-score on the benchmark datasets, outperforming state-of-the-art methods. Future research will explore the application of SpeechHGT to other neurodegenerative diseases and datasets, alongside architectural enhancements to further refine its diagnostic capabilities. These findings underscore the potential of hypergraph-based learning frameworks to advance noninvasive, speech-based diagnostic tools, providing new insights into the cognitive decline associated with Alzheimer's disease.

Acknowledgments

This work is supported by the National Language Commission Key Program (No. ZDI145-80) and the Natural Science Foundation of Liaoning (Grant No. 2024-MSLH-073).