Foundation models are widely used for compressing complex data into vector embeddings (vembs), offering reduced storage and computational efficiency, and significant improvements in diagnostic accuracy, automation, and efficiency in medical imaging. However, concerns remain that these vembs may encode demographic features, which could provide a pathway for bias in AI models in medical imaging. This study investigates whether demographic attributes - including sex, age, ethnicity, and insurance type - are embedded in vembs derived from chest radiographs in the MIMIC-CXR and CheXpert datasets. We generate vembs using three different state-of-the-art contrastive learning-based foundation models, namely, CXR Foundation, MedCLIP, and BiomedCLIP, assessing demographic predictability. Through rigorous statistical analysis and machine learning evaluations, we demonstrate substantial demographic encoding, indicating a plausible pathway through which bias may propagate. Our findings provide cautionary evidence supporting the need for further investigation and auditing of potential biases in vemb-based medical imaging predictions.
Voice-based screening offers a noninvasive, scalable avenue for early detection of type 2 diabetes using everyday smartphone recordings and acoustic features alone. We present V2D (Voice2Diabetes), a novel application of spectrogram-transformer embeddings derived exclusively from short speech segments for patient-level diabetes classification, without requiring any clinical measures or demographic variables. Adults (n=461; 157 female, 304 male) each completed multiple smartphone recordings while reading randomly selected sentences on their own smartphones. Mel-spectrograms were encoded with a pretrained Audio Spectrogram Transformer (AST) to train sex-stratified patient-level classifiers under fivefold nested cross-validation with a held-out calibration set; predictions were aggregated across recordings per participant. Using acoustic features alone, the models achieved patient-level balanced accuracy of 0.724 ± 0.023 (males) and 0.713 ± 0.021 (females), with area under the receiver operating characteristic curve (AUC) of 0.779 ± 0.019 and 0.788 ± 0.018, respectively, averaged over five independent random seeds. The pipeline incorporated probability calibration, sensitivity-first threshold optimization, and optional interpretable acoustic anchors (e.g., fundamental frequency, harmonic-to-noise ratio) to support clinical interpretation. These results provide a rigorous technical validation of AST-derived embeddings for acoustic-only, sex-stratified T2D classification in smartphone recordings and motivate prospective external validation in broader populations.
Sepsis is a life-threatening condition that can be fatal. Advances in big data analytics and the data-rich environment of intensive care units have enabled the development of artificial intelligence-based early warning systems that offer a solution to reduce sepsis-related mortality. Support Vector Machines (SVMs) establish optimal classifiers and deep learning automatically learns key features. This paper proposes an early warning system for sepsis that combines a SVM and deep learning. A hybrid encoder model that integrates a convolutional neural network (CNN), a bi-directional long short-term memory (BiLSTM), an attention mechanism (AM) and a fully connected network (FCN) are used to extract critical features. To improve feature robustness, a two-channel self-regulated generative adversarial network is used: the cooperative channel learns predictive features and the adversarial channel prevents the acquisition of erroneous features. The refined features are then input into a novel deep transfer least squares fuzzy hyperplane-based support vector machine (DT-LS FH SVM) that uses statistical learning theory, fuzzy theory, and transfer learning to construct an optimal sepsis warning system. This warning system predicts sepsis-related adverse outcomes up to 28 days in advance to give early warning and mitigate the risk of septic-related mortality. The proposed method demonstrates satisfactory predictive performance, with an area under the receiver operating characteristic curve (AUROC) of 0.793589, highlighting the feasibility of leveraging early integrated data to predict adverse outcomes within 28 days among patients with sepsis. The experimental results show that the proposed model outperforms current state-of-the-art methods.
We propose an integrated gaze-driven framework for assistive robotic manipulation that infers user intent to support fluid human-robot collaboration in semi-structured manipulation environments. The system consists of three functional components: (I) a Gaze-based Teleoperation method using Gaussian Mixture Regression (GMR) to generate smooth, intent-aligned manipulator motion; (II) a Gaze-Driven Grasp Pose Detection (GD-GPD) algorithm that fuses geometric grasp quality with gaze-derived attention using an arbitration factor ($\alpha$) to select intent-aligned grasp poses; and (III) a Gaze Guidance Pose Optimization strategy that refines placement position and release orientation to improve placement stability. In a user study with ten participants, the proposed Intent-Inference Control (IIC) mode was compared with both a Gaze-Triggered Control (GTC) mode and a Traditional Pipeline Control (TPC) mode. The results showed that IIC achieved higher target grasp accuracy and shorter completion time than GTC, demonstrating the benefit of gaze-weighted grasp candidate ranking over fixation-based target selection. Compared with TPC, IIC also improved task success rate, completion time, target grasp accuracy, and placement stability, confirming the advantage of integrating gaze-based motion guidance, intent-aware grasp selection, and placement optimization. Overall, the proposed framework provides a practical step toward more natural and responsive assistive robotic manipulation across the motion-grasp-placement sequence.
Background and Motivation: Classification of transcriptomic gene data is essential for Cardiovascular disease (CVD) risk, particularly in Hypertrophic Cardiomyopathy (HCM) and Acute Myocardial Infarction (AMI) patients. Existing approaches suffer from limited feature representation and weak biological context modeling. To address these gaps, we propose AtheroEdge™ 5.0, which incorporates three novel Transformers (Xmers): Neuro-Topology (NT), Self-Supervised Contrastive Learning with Alignment and Random Feature Masking (SCARF), and Temporal Diffusion Gene (TDG). Method: Twelve Artificial Intelligence models: three novel Xmers (Models A), three Legacy Xmers (Models B), three Deep Learning (Models C), and three machine learning models (Models D) were designed. Feature engineering included Differential Expression Analysis (DEA) for gene selection and normalization of two different cardiac datasets: HCM and AMI. (iii) Performance was evaluated using K10 cross-validation. The AtheroEdge™ 5.0 was scientifically validated using (a) unseen datasets, (b) K-effect, (c) Generalization-effect, and (d) Local Interpretable Model-agnostic Explanations (LIME)-based Models. Software verification was conducted using Coronary Artery Disease data. Reliability and stability tests were conducted. We hypothesized that: (a) Models A outperform Models B to D, (b) unseen data performance is comparable to seen data for both HCM and AMI datasets, and (c) TDG-Xmer outperforms NT Xmer and SCARF-Xmer. Results: Model A achieved a mean accuracy superior to Models B, C, and D by 4.01%, 10%, and 23.95%, respectively. The Mean Area-under-the-curve of Models A, B, C, and D were 0.96, 0.95, 0.91, and 0.80, respectively. Performance decline on unseen cohorts remained below 10%, meeting regulatory criteria. 87% of high-risk genes were consistently identified by all three novel Xmers and by DEA. TDG-Xmer outperformed NT-Xmer and SCARF-Xmer by 0.5% and 5.26%, respectively. Conclusions: The proposed Xmers provide a robust and scientifically validated framework for accurate CVD risk stratification.
Self-supervised monocular depth estimation in endoscopy is fundamentally constrained by the ill-posed nature of photometric supervision. In this work, we identify a critical yet overlooked cause of this ambiguity: the inherent insensitivity of photometric loss to depth noise. To overcome this intrinsic limitation, we propose Depth Error Calibration Learning (DECL), a two-stage framework that suppresses prediction variance and mitigates residual errors in self-supervised depth estimation. In Stage I (Variance Reduction), a cyclic depth generation strategy produces multiple depth hypotheses for the input image. The per-pixel empirical variance is quantified and integrated into a dedicated variance loss term, which penalizes inconsistent predictions and encourages the network to generate more stable and reliable depth estimates. In Stage II (Bias Calibration), an image-conditioned diffusion model refines the Stage-I depth prior and mitigates structured residuals through iterative denoising, thereby improving geometric accuracy and global consistency. Extensive experiments on three public endoscopic datasets demonstrate that DECL achieves consistent improvements over representative self-supervised monocular depth estimation methods under the evaluated protocols. Moreover, ablation studies on two representative backbones indicate that DECL is not restricted to a single network implementation, while broader validation on additional backbone families remains necessary. The source code is publicly available at https://github.com/DavidLuBit/EndoDenoising.
Automatic skin lesion segmentation is a pivotal problem in the medical domain and an indispensable component in the computer-aided diagnosis program. Most convolutional neural network-based segmentation algorithms have demonstrated promising performance due to their ability to encode detail and semantic features efficiently. However, they fail to capture the long-range contextual information at the global level. Therefore, researchers employ Transformer architecture to address this issue. Unfortunately, these methods fail to learn sufficient pixel information at the local level. Motivated by this, some researchers attempt to design a hybrid architecture based on CNN and Transformer. However, the large number of parameters and high computational cost make them challenging to train and use. To alleviate these problems, we propose an effective Lightweight Multi-scale Deformation Aggregation U-Net (LMDAU-Net), which consists of a Lightweight Local-global Learning Module (LLM) and an Adaptive Interactive Fusion Module (AIF). Specifically, we utilize the two branches of the proposed LLM to efficiently learn local fine-grained and global coarse-grained features that assist the model in capturing the complementary feature representations. Moreover, we employ the AIF to selectively learn semantic and detail features at different scales, which can dynamically explore variable feature cues. Extensive experiments on four skin benchmarks, including ISIC2016, ISIC 2017, ISIC2018, and PH2, demonstrate that LMDAU-Net achieves state-of-the-art performance in both qualitative and quantitative aspects. We have released our code on https://github.com/Lm0611/LMDAU-Net.
Polyp segmentation in colonoscopy images plays a pivotal role in computer-aided medical diagnosis and the early prevention of colorectal cancer. However, existing methods often suffer from performance degradation when confronted with extreme polyp scale variation and polyp boundary ambiguity. To address these challenges, we propose the Staged Global-to-Local Cross-Scale Fusion Network (SGLF-Net), which adopts a novel staged global-to-local learning paradigm to progressively refine segmentation from coarse global semantics to fine-grained local details. Specifically, the Global Semantic Perception Stage integrates a Swin Transformer Encoder and a Dynamic Attentive Decoder (DAD) to construct comprehensive multi-scale contextual representations. The Local Detail Refinement Stage employs an Edge-aware Dynamic Attentive Decoder (E-DAD) to enhance structural fidelity and boundary precision through explicit edge-guided supervision. Furthermore, we introduce the Cross Spatial-Scale Feature Aggregation and Reconstitution (CSSAR) module, equipped with hybrid attention mechanisms, to facilitate efficient semantic structural interaction between the two cascaded stages. Extensive experiments on five public benchmark datasets demonstrate that SGLF-Net consistently outperforms state-of-the-art methods in both segmentation accuracy and boundary preservation.
Stroke is one of the leading causes of long-term motor disability worldwide, placing a substantial burden on individuals, families, and healthcare systems. Innovative rehabilitation strategies such as motor imagery-based brain-computer interface (MI-BCI) are critical to accelerating stroke recovery. However, current MI-BCI methods face key challenges: low generalizability due to cross-patient variability, lack of effective functional assessment, limited availability of patient data, coupled with the lack of effective data augmentation approaches. To address these issues, we propose a unified EEG-based framework that simultaneously performs motor imagery classification, hemiplegic side detection, and functional recovery prediction. Our method introduces a diffusion model tailored to the spatio-temporal characteristics of EEG, incorporating a decoupled neural architecture with rotary spatial encoding and autoregressive temporal fusion. To mitigate data scarcity, we design two augmentation strategies specifically adapted to the characteristics of stroke EEG. Extensive experiments demonstrate superior performance and generalizability across multiple MI-BCI tasks, supporting the potential of the method for deployment in personalized stroke rehabilitation.
Accurate sleep staging is essential for the diagnosis and evaluation of sleep disorders. However, convolutional neural network (CNN)-based methods are limited in capturing stage-specific sleep waveforms, and classification errors often occur in transitional sleep regions. In addition, existing models do not adequately account for inter-individual variability related to physiological factors such as age and sex. To address these issues, we propose ViSTA-SleepNet, a two-stage, multi-view, multimodal framework for sleep staging.Specifically, data augmentation is applied to sleep stage transition segments to improve boundary learning. For EEG signals, a multi-view representation learning strategy and a cross-view Transformer are used to jointly model raw waveforms, amplitude variations, and temporally structured events, thereby enhancing the detection of key sleep events. EOG and EMG are modeled as complementary modalities and adaptively fused with EEG features. In addition, a feature-wise linear modulation (FiLM) mechanism based on demographic and physiological information is introduced to support individualized representation learning and joint detection of sleep stages, spindles, and slow waves. Finally, a bidirectional gated recurrent unit (Bi-GRU) and conditional random field (CRF) are combined to improve the physiological consistency of sleep stage transitions.Experiments on Sleep-EDF-20, Sleep-EDF-78, and SHHS using strict 20-fold subject-independent cross-validation show that ViSTA-SleepNet achieves accuracies of 88.5%, 84.5%, and 89.6% for sleep staging, respectively, while exceeding 95% accuracy in spindle and slow-wave detection. Visualization results further confirm its advantages in discriminative performance and physiological plausibility.
Idiopathic scoliosis (IS) is a prevalent spinal deformity that impairs posture and psychosocial health. Physiotherapeutic scoliosis-specific exercise (PSSE) is an effective conservative treatment that needs long-term and regular training. However, academic demands, geographic barriers, and time constraints limit frequent clinic visits, making home-based PSSE the mainstay of care. Also, the lack of supervision makes correct execution and sustained adherence difficult, posing a major challenge in spinal rehabilitation. To address this issue, we innovatively equate the spine as an open-chain spatial multibody kinematic mechanism and propose an IMU-driven, kinematics- and CT-integrated dynamic spine model (KCT-DSM). By fusing real-time motion data with CT-derived anatomical information, KCT-DSM enables continuous tracking of spinal kinematics. Further, we develop a home-based PSSE rehabilitation training system integrating KCT-DSM and PSSE. A preliminary feasibility evaluation was conducted in patients with IS. The Cobb-like angle derived from KCT-DSM differs from X-ray measurements by 1.56$^{\circ }$, suggesting that KCT-DSM provides motion tracking of spinal deformity with different scoliosis patterns. In addition, when using the training system, the effective duration of standard rehabilitation exercises increased by 45.7% in the pilot cases. Overall, the proposed system has potential for providing quantitative feedback and supporting home-based PSSE training.
Cryptic binding sites (CBSs) are crucial functional sites that become accessible following conformational changes by ligand binding. They play a significant role in expanding the scope of druggable targets and revealing the dynamic regulatory mechanisms of proteins. Most existing computational methods rely on holo (ligand-bound) structures and struggle to effectively identify CBSs in the apo (ligand-free) state. Furthermore, these methods fail to account for protein three-dimensional conformational changes and spatial geometric information, often making it difficult to explain the dynamic characteristics of CBS formation.To address these challenges, we present CrypKANet, an innovative multimodal predictive framework for cryptic binding site identification that integrates EGNN, gated attention mechanisms, and Kolmogorov-Arnold Network (KAN). Our framework employs a two-branch design to independently encode geometric and biochemical features: the EGNN branch is dedicated to explicitly capturing three-dimensional spatial restraints and conformational dynamics, whereas the GINE branch strengthens the representation of residue-level chemical interactions and topological connectivity. Finally, the model predicts cryptic binding sites through the KAN classifier. Experimental results on the CBS benchmark dataset demonstrate that CrypKANet outperforms the existing state-of-the-art methods, and exhibits excellent generalization performance on protein-protein interaction sites and ligand binding site tasks.
Corneal endothelium segmentation is crucial for clinical evaluation, but remains constrained by two significant challenges: the scarcity of high-quality annotations and the difficulty of reliably delineating complex pathological patterns. In particular, the corneal guttata exhibits high structural variability, resulting in limited and inconsistent training data. Previous studies rely heavily on manual labeling or GAN-based synthesis, which cannot generate text-controlled paired image-mask datasets at scale and lack the required fidelity for pretraining modern segmentation models. To overcome these limitations, this study introduces a two-stage training framework that integrates text-guided diffusion models, transfer learning, and boundary-refinement algorithms. A Stable Diffusion model is fine-tuned with LoRA and conditioned with ControlNet to generate three categories of synthetic data: endothelial images with guttata, without guttata, and their corresponding binary boundary masks. U-Net, TransUnet, and SwinUnet are then pretrained on synthetic paired datasets and fine-tuned on limited real annotations, with a watershed algorithm incorporated as a post-processing step to enhance boundary continuity in the final segmentation outputs. Experimental results demonstrate that increasing the size of synthetic pretraining data generally improves segmentation accuracy. Among the models evaluated, U-Net achieves the highest segmentation performance, indicating that local feature extraction is particularly well suited to corneal endothelial morphology compared with global self-attention mechanisms. Although watershed post-processing may slightly lower Dice scores due to its sensitivity to overlap metrics, it offers a clinically meaningful trade-off by substantially improving boundary continuity and yielding more anatomically coherent cell borders.
Enzyme optimal pH is a key determinant of catalytic activity, yet existing computational methods often treat it as a fixed value and rely mainly on sequence features. In reality, enzyme activity is stable over a pH range and is strongly influenced by three-dimensional structural context. Structural information is underutilized due to limited experimental data. To address this gap, we present DeepPH, a structure-aware framework that models enzyme optimal pH as an interval regression problem to capture inherent uncertainty. DeepPH integrates sequence embeddings with residue-level features from predicted protein structures and encodes three-dimensional geometry using spatial radius graphs and E(3)-equivariant message-passing networks. An attention mechanism adaptively fuses biochemical and structural information. We further conducted downstream analyses, including residue-level attention, solvent accessibility, and three-dimensional visualization, to uncover the biochemical and structural determinants learned by the model. Case studies demonstrate that interval predictions more accurately reflect enzymes with broad pH activity profiles. Extensive experiments show that DeepPH outperforms existing methods under both standard and interval-aware evaluations and generalizes well to extreme-length sequences. The code, datasets, and supplementary materials are publicly available at https://github.com/LabJunBMI/DeepPH.git.
Accurate localization of the seizure onset zone (SOZ) is essential for the surgical treatment of drug-resistant epilepsy, yet it remains challenging because seizures involve rapid and variable spatiotemporal reorganization of brain networks. In this study, we propose a Koopman-inspired dynamic graph (KDG) framework for SOZ localization from intracranial electroen cephalography (iEEG). Unlike conventional high-frequency oscillation or functional connectivity (FC) approaches that primarily characterize local events or isolated network snapshots, the proposed framework explicitly models the temporal evolution of FC graphs using an operator-based dynamical representation. The estimated evolution operator is decomposed using singular value decomposition (SVD), and the dominant singular components are mapped back to graph space to derive interpretable nodal graph descriptors, including degree variation, node strength, eigenvector centrality, and nodal global efficiency. Under leave one-subject-out (LOSO) validation in seizure-free patients, The proposed KDG features achieved an area under the receiver operating characteristic curve (AUC) of 0.84, accuracy of 0.76, sensitivity of 0.79, and specificity of 0.76 for SOZ localization. Feature importance analysis showed that earlyictal periods contributed most strongly to SOZ discrimination, suggesting that early seizure-related network reorganization contains clinically informative biomarkers for epileptogenic tissue localization. These findings demonstrate that modeling dynamic functional brain networks can provide interpretable and clinically relevant biomarkers for SOZ localization and may offer a general framework for characterizing time-varying pathological network dynamics.
Sparse-view CT reconstruction aims to synthesize volumetric CT images from a limited number of X-ray projections, reducing radiation dose while maintaining diagnostic quality. However, the substantial cross-modal gap between 2D X-rays and 3D CT volumes presents significant challenges, including uneven distribution of information across different views, artifact issues, and excessive computational costs. Therefore, in this paper, we propose MAC-DiffCT, a multi-scale adaptive conditional diffusion model designed for accurate and efficient CT reconstruction from biplanar X-rays. Our approach first extracts multi-scale 2D features from multi-view X-rays using a UNet-based encoder. A novel Bi-Directional Cross-Attention (BDC-Att) module adaptively fuses features by assigning spatially varying weights to each view. We then introduce a Multi-Scale Feature Sampling (MS-FS) module that projects 3D coordinates onto 2D planes, sample features across scales, and integrates them via a multi-layer perceptron to form a latent structural representation. This 3D structural feature serves as a condition for a latent-space conditional diffusion model, which reconstructs high-quality CT volumes with enhanced anatomical fidelity. An additional signed distance function (SDF) loss is applied to promote structural consistency in the 3D space. Experimental results on both public and private chest datasets demonstrate that MAC-DiffCT consistently outperforms existing methods, achieving the highest reconstruction accuracy with PSNR of 27.68 and 25.43 dB, SSIM of 0.8223 and 0.7598, and the lowest LPIPS of 0.0992 and 0.1322. Downstream evaluation via lung segmentation and an interpretability study further highlight the transparent, explainable, and anatomically grounded nature of our model. MAC-DiffCT offers a low-radiation alternative to conventional CT, especially for vulnerable patients requiring repeated imaging and intraoperative scenarios where CT use is constrained.
Dynamic MRI suffers from limited spatiotemporal resolution due to long acquisition times. Undersam pling k-space can accelerate imaging, but it poses a major challenge for accurate image reconstruction. Supervised deep learning methods achieve impressive results, yet they rely on large fully sampled datasets which are often difficult to obtain. Recently, implicit neural representations (INR) have emerged as a powerful unsupervised paradigm that reconstructs images from a single undersampled dataset without external training data. However, existing INR-based methods still face challenges when applied to highly under sampled dynamic MRI, mainly due to their inefficient representation capacity and high computational cost. To address these issues, we propose TenF-INR, a novel unsupervised framework that integrates low-rank tensor modeling with INR, where each factor matrix in the tensor decomposition is modeled as a learnable factor function. Specifically, we employ INR to model learnable tensor functions within a low-rank decomposition, reducing the parameter space and computational burden. A patch-based nonlocal tensor modeling strategy further exploits temporal correlations and inter-patch similarities, enhancing the recovery of fine spatiotemporal details. Experiments on dynamic cardiac and abdominal datasets demonstrate that TenF INR achieves up to 21-fold acceleration, outperforming both supervised and unsupervised state-of-the-art methods in image quality, temporal fidelity, and quantitative accuracy. This work establishes an unsupervised learning framework for high-dimensional dynamic MRI reconstruction, offering a promising direction for efficient and interpretable modeling in advanced medical imaging applications.
Psychotherapy summarisation poses unique challenges due to the emotional nuance, relational complexity, and non-linear discourse of therapeutic dialogues. Existing large language models (LLMs), even when fine-tuned, often fail to capture these structures, leading to clinically incomplete or incoherent summaries. In this work, we propose TheraFuse, a clinically grounded summarisation framework that adapts discourse-level relational graphs and LLaMA fusion to the psychotherapy domain. Our method first parses therapy dialogues into discourse graphs and encodes them using Relational Graph Convolutional Networks. These representations are then injected into LLaMA's hidden layers to guide generation. Experiments on a benchmark counselling dataset (iHOPE), following a benchmark counselling summary template (iCARE), show that TheraFuse significantly outperforms both zero-shot and fine-tuned baselines in both surface metrics (BLEU, ROUGE) and semantic alignment (BERTScore, InfoLM); these gains are statistically significant under paired $t$-tests with Holm-Bonferroni correction ($p< 0.05$, Section VI-A.2). Human evaluation by mental health professionals confirms the improved coherence, structure, and clinical usefulness of the generated summaries.
Malpositioned medical tubes and catheters are a persistent source of iatrogenic harm in critical care, yet the radiographic interpretation burden on intensive care units remains substantial. We present ESCA-Net, a knowledge-guided dual-task deep learning framework for simultaneous multi-label classification of endotracheal tubes (ETT), nasogastric tubes (NGT), central venous catheters (CVC), and Swan-Ganz catheters in portable chest radiographs (CXRs). The architecture couples an EfficientNet-Lite4 encoder with a UPerNet segmentation decoder and a spatial attention classification head, jointly optimised via focal loss and a hybrid segmentation loss to address extreme class imbalance. Clinical placement knowledge is encoded as structured pixel-level supervision, instantiating knowledge-guided multimodal fusion in which heterogeneous annotation modalities jointly constrain encoder representations toward clinically grounded features. The framework achieves a mean AUC-ROC of $0.9618 \pm 0.0016$ across five-fold cross-validation, with Swan-Ganz detection at AUC 0.9980 and ETT categories at mean AUC 0.9840. Auxiliary segmentation supervision yields targeted gains in CVC classification ($+1.45\%$ AUC) and CVC-Borderline ($+1.89\%$), demonstrating that approximate pixel-level anatomical knowledge provides effective spatial regularisation for borderline classification under severe class imbalance, without requiring formal knowledge graph construction at inference time.
Non-invasive diagnostic technologies for type 2 diabetes mellitus (T2DM) have emerged to address the limitations inherent in conventional invasive detection methods. Blood glucose levels may modulate laryngeal neuromuscular activity, thereby supporting the investigation of voice-based T2DM detection as a potential approach. However, methodological constraints in data acquisition and analytical limitations in extant studies have impeded substantive progress in voice-based T2DM detection. To systematically investigate voice-based T2DM detection, we conducted a comprehensive acoustic analysis of 408 participants, comprising both non-diabetic (ND) controls and T2DM patients. Specifically, we collected sustained vowel phonations and standardized short sentences from each participant to develop the multi-scale cooperative vowel attention strategy (MsCVA). This architecture integrates two complementary modules: the vowel stream analysis (VSA) module and the speech stream analysis (SSA) module to extract discriminative feature representations. Through multimodal feature fusion, MsCVA achieved an accuracy of 80.12% in voice-based T2DM detection.
