IEEE Transactions on Biomedical Engineering最新文献

The hybrid EEG-fNIRS Brain-computer interface (BCI) combines the high temporal resolution of electroencephalography (EEG) with the high spatial resolution of functional near-infrared spectroscopy (fNIRS) to enable comprehensive brain activity detection. However, integrating these modalities to obtain highly discriminative features remains challenging. Most existing methods fail to effectively capture the spatiotemporal coupling features and correlations between EEG and fNIRS signals. Furthermore, these methods adopt a holistic learning paradigm for the representation of each modality, leading to unrefined and redundant multimodal representations. To address these challenges, we propose a disentangled multimodal spatiotemporal learning (DMSL) method for hybrid EEG-fNIRS BCI systems, which simultaneously performs multimodal spatiotemporal coupling and disentangled representation learning within a unified framework. Specifically, DMSL utilizes a compact convolutional module with one-dimensional temporal and spatial convolution layers to extract complex spatiotemporal patterns from each modality and introduces a multimodal attention interaction module to comprehensively capture the inter-modality correlations, enhancing the representations for each modality. Subsequently, DMSL designs an adaptive multi-branch graph convolutional module based on reconstructed channels to effectively capture the spatiotemporal coupling features, incorporating modality consistency and disparity constraints to disentangle common and modality-specific representations for each modality. These disentangled representations are finally adaptively fused to perform different task predictions. The proposed DMSL demonstrates state-of-the-art performance on publicly available datasets for mental arithmetic, motor imagery, and emotion recognition tasks, exceeding the best baselines by 2.34%, 0.59%, and 1.47%, respectively. These results demonstrate the effectiveness of DMSL in improving EEG-fNIRS decoding and its strong generalization ability in BCI applications.

Gait asymmetry is a significant clinical characteristic of hemiplegic gait that most stroke survivors suffer, leading to limited mobility and long-term negative impacts on their quality of life. Although a variety of exoskeleton controls have been developed for robot-aided gait rehabilitation, little attention has been paid to correcting the gait asymmetry of stroke patients, and it remains challenging to properly share control between the exoskeleton and patients with partial motor control. In view of this, an assist-as-needed (AAN) hip exoskeleton control with human-in-the-loop optimization is proposed to correct gait asymmetry in hemiplegic gait. To realize the AAN concept, an objective function was designed for real-time evaluation of the subject's gait performance and active participation, which considers the variability of natural human movement and guides the online tuning of control parameters on a subject-specific basis. In this way, subjects were stimulated to contribute as much as possible to movement, thus maximizing the efficiency and outcomes of gait rehabilitation. Finally, an experimental study was conducted to verify the feasibility of the proposed control with simulated hemiplegic gait, and the common hypothesis that AAN controls can improve active human participation was clearly validated from a biomechanics perspective.

Objective: Accurate segmentation of heart sound signal stages is critical in cardiovascular disease analysis.

Methods: This study proposed the integration of a duration hidden Markov model (DHMM) with a temporal convolutional network (TCN) and an adaptive calibration mechanism (based on electrocardiogram signals) to enable the precise segmentation of complex heart sound signals. Multiple features of heart sound signals are extracted and utilized as model inputs, constructed a segmentation model architecture improved by TCN-based observation probability estimation and an attention mechanism integrated into the Viterbi algorithm.

Results: The experimental results demonstrated that the average accuracy of this method is 94.71 ± 2.64% at a segmentation error of 50ms. The enhanced Viterbi algorithm elevated performance by approximately 9 percentage points. Furthermore, the adaptive calibration mechanism yielded an additional average accuracy increase of 1.41 percentage points and reduced the standard deviation by 1.21 percentage points. Conclusion: Compared to traditional methods employing Gaussian distribution-based observation probability estimation, the utilization of a TCN substantially enhanced state discrimination accuracy, achieving an improvement of approximately 3 percentage points. The refined Viterbi algorithm demonstrated superior performance relative to prior methodologies.

Significance: This method enables effective segmentation of complex heart sound data, delivering a high-precision solution for the automated analysis of heart sounds. Our code can be found in https://github.com/KC-Y-bjut/Heart-sound-segmentation.

Objective: Harmonic Motion Imaging (HMI) is an ultrasound elasticity imaging method that measures the mechanical properties of tissue using amplitude-modulated acoustic radiation force (AM-ARF). Multi-frequency HMI (MF-HMI) excites tissue at various AM frequencies simultaneously, allowing for image optimization without prior knowledge of inclusion size and stiffness. However, challenges remain in size estimation as inconsistent boundary effects result in different perceived sizes across AM frequencies. Herein, we developed an automated assessment method for tumor and focused ultrasound surgery (FUS) induced lesions using a transformer-based multi-modality neural network, HMINet, and further automated neoadjuvant chemotherapy (NACT) response prediction. HMINet was trained on 380 pairs of MF-HMI and B-mode images of phantoms and in vivo orthotopic breast cancer mice (4T1). Test datasets included phantoms (n = 32), in vivo 4T1 mice (n = 24), breast cancer patients (n = 20), FUS-induced lesions in ex vivo animal tissue and in vivo clinical settings with real-time inference, with average segmentation accuracy (Dice) of 0.91, 0.83, 0.80, and 0.81, respectively. HMINet outperformed state-of-the-art models; we also demonstrated the enhanced robustness of the multi-modality strategy over B-mode-only, both quantitatively through Dice scores and in terms of interpretation using saliency analysis. The contribution of AM frequency based on the number of salient pixels showed that the most significant AM frequencies are 800 and 200 Hz across clinical cases.

Significance: We developed an automated, multimodality ultrasound-based tumor and FUS lesion assessment method, which facilitates the clinical translation of stiffness-based breast cancer treatment response prediction and real-time image-guided FUS therapy.

Objective: Real-time estimation of brain state is essential for efficient brain stimulation. Specifically, the electroencephalography (EEG) oscillation phase arose as a promising biomarker for instantaneous brain excitability, making it ideal for state-dependent brain stimulation. Current methods for real-time EEG phase extraction lose accuracy in the presence of non-stationary noise, motivating the development of a more robust and accurate algorithm. Here, we propose and validate Bayesian Temporal Prediction (BTP) as an effective method for EEG phase detection in real-time.

Methods: BTP utilizes a short pre-session EEG recording and learning of the personalized prediction parameters, enabling subsequent high-precision real-time phase detection. We experimentally validate BTP in humans and compare its performance to a strong benchmark algorithm.

Results: BTP demonstrates accurate EEG oscillation phase detection across a broad range of conditions and target oscillations, facilitating personalized brain stimulation.

Conclusion: This study introduces BTP as a robust, computationally efficient, and accurate method for EEG state-dependent stimulation.

Significance: The widespread adoption of BTP in research and clinical settings has the potential to enhance treatment efficacy and minimize inter- and intra-individual variability in brain stimulation interventions.

Objective: Data augmentation is important for enhancing subject-independent classification in deep learning (DL) approaches for steady-state visual evoked potential (SSVEP) brain-computer interfaces (BCIs) using electroencephalography (EEG). However, current augmentation techniques often inadequately exploit individual-specific style characteristics, limiting the model's robustness against inter-subject style variability. To tackle this problem, this study proposes a novel data augmentation method called Simultaneous Spatial-Energy Representation (SSER).

Methods: SSER employs singular value decomposition (SVD) to extract spatial and energy representations from EEG signals, effectively capturing style characteristics. These representations are independently mixed across source domains during signal reconstruction, generating novel domains that cover a broader range of styles. This strategy promotes the learning of domain-invariant features and enhances the model's robustness to style variability.

Results: Comprehensive experiments on public datasets demonstrate that SSER outperforms state-of-the-art data augmentation techniques and generalizes well across various DL models. Furthermore, self-collected offline and online experiments involving 30 subjects provide additional evidence of the method's effectiveness.

Conclusion: By simultaneously manipulating spatial and energy representations, SSER offers a richer characterization of EEG signal style variability, leading to superior performance.

Significance: The proposed innovative data augmentation method advances subject-independent classification, facilitating the broader application of EEG-based BCIs in real-world scenarios.

Objective: Stress has emerged as a major adverse factor affecting both physical and mental health. Timely recognition of stress is essential for effective stress management. However, substantial inter-individual variability in stress responses often leads to distribution shifts, thereby making conventional generic stress recognition models ineffective in adapting to such changes.

Methods: To address distribution shifts caused by inter-individual variability, this paper proposes the Variational Instance-Adaptive Stress recognition (VIAStress), a Domain Generalization (DG)-based framework that consists of two main components: Variational Instance Adaptation (VIA) and Multimodal Domain Generalization (MMDG). The VIA leverages the distribution-fitting capability of variational inference to model classifier parameters as instance-conditioned distributions, dynamically adjusting decision boundaries for personalized stress recognition. In addition, the model's feature extractor integrates the MMDG strategy to facilitate information interaction between PPG and EDA, thereby enhancing the generalization of multimodal feature representations.

Results: VIAStress is evaluated on four public datasets, WESAD, UBFC-Phys, VerBIO, and CAN-STRESS. Under both within-dataset and cross-dataset subject-independent settings, the proposed method achieves superior generalization performance on unseen subjects in most tasks, outperforming competitive approaches.

Conclusion and significance: This study introduces VIAStress, a personalized stress recognition model designed to address the practical challenges of inter-individual variability. The results show that VIAStress exhibits strong adaptability and robustness to user differences, supporting the development of more effective and personalized stress management solutions.

Background: Bladder cancer (BC) symptoms often overlap with benign conditions, while no routine screening exists for general population. We aim to develop a machine learning (ML)-based screening pipeline for early BC detection using electronic-health-records (EHRs) in primary care.

Methods: A multi-centred case-control cohort (1995-2018; n = 64,884) was created for model training and testing. We further validated the model prospectively on an independent cohort (2019-2020; n = 4,569). We proposed the Parsimony driven REweighting for Calibrated Input-based Screening for Early detection-Adjustable Grey Zone (PRECISE-AGZ), which identified influential features from 48,261 candidates and developed a calibrated logistic regression screening model with optimised grey-zone thresholds.

Results: We finally identified 38 features, achieving an AUC (area under the curve) of 0.789 (95% CI: 0.780-0.798) on testing set. Neurological disorders (e.g., Parkinson's disease, OR: 0.86, 95% CI: 0.79-0.92) and medications (e.g., Tamoxifen, OR: 1.13, 95% CI: 1.07-1.20) emerged as novel predictors for BC screening. The screening model stratified the population into three risk categories based on predicted probability: low-risk (0.55), achieving a sensitivity of 0.852, F1-score of 0.799, and screening population coverage (SPC) of 34.5%. Applied to the prospective validation cohort, model performance varied by months before BC diagnosis, with sensitivities ranging from 0.872 (F1-score: 0.714, SPC: 29.9%) at the first month to 0.667 (F1-score: 0.690, SPC: 12.7%) at the twelfth month.

Conclusion: The PRECISE-AGZ pipeline efficiently identified clinical signals from EHRs for early BC detection, offering promising potential for implementing population-based BC screening.

Objective: Deep brain stimulation (DBS) is increasingly used in treating motor disorders, neurodegenerative conditions, and mental health conditions. However, treatments are limited to the most severe cases due to the invasiveness of the surgical procedure. We propose stimulating ventral brain areas using electrodes inserted through the nose and placed in contact with the inferior skull base bones, under the olfactory cleft and in the sphenoid sinus.

Methods: We designed a compact, low-power stimulation implant that fits in the sphenoid sinus and tested the implant insertion and operation in cadaveric heads, measuring the electric field obtained in deep brain regions. Using commercial multielectrode probes and an external stimulator, we also measured the electric field in cadaveric brains generated by different electrode placements to characterize the focus and steering that can be achieved.

Results: The implant can generate electric fields in the brain above 10 V/m, likely sufficient for neuronal activation. It is powered using either a 3 V battery which can generate almost 100 000 pulses from a 1 mAh charge, or alternatively, using inductive wireless power transfer, with the primary coil placed around the circumference of the head.

Conclusion: This first-of-its-kind sphenoid implant demonstrates the possibility of minimally-invasive, focused deep brain stimulation.

Significance: By enabling stimulation of deep brain regions without requiring surgery, transnasal stimulation can drastically improve the accessibility of some DBS treatments and broaden its applicability to additional mental health conditions.

Objective: To reduce variability in chest radiography (CXR) from acquisition and post-processing, and assess whether harmonization improves image quality and down-stream diagnostic performance.

Methods: A generative adversarial network (GAN) was trained exclusively on virtual images produced by a Virtual Imaging Trial (VIT). The model maps non-harmonized CXRs to noise-free, unprocessed references using a dual U-Net with contrastive loss. Evaluation spanned virtual, physical-phantom, and clinical data using generic (GIQMs) and chest-specific (CIQMs) metrics. A phantom benchmark compared the method with ComBat harmonization and a conventional denoising algorithm. Clinical generalization was tested on VinDr-CXRs, examining feature compactness via Uniform Manifold Approximation and Projection (UMAP). We also assessed NIH-CXRs, quantifying diagnostic accuracy with bootstrap uncertainty.

Results: Compared with non-harmonized images, harmonized CXRs yielded approximately 89% lower NRMSE, 50% higher PSNR, and 86% higher SSIM. On phantom data, CIQM variability fell from 0.255 to 0.026 and was reduced more consistently than with ComBat or denoising. Clinical analyses on VinDr-CXRs showed tighter UMAP clusters for harmonized features, indicating suppression of acquisition-related variability. On NIH-CXRs, training and testing a classifier in the harmonized domain improved diagnostic accuracy over the non-harmonized domain and lowered cross-domain sensitivity; pleural and cardiac categories showed consistent gains, while texture-dependent labels exhibited task-dependent effects.

Conclusion: VIT-guided training enables a GAN to harmonize CXRs toward a raw reference, improving quantification variability and harmonizing image representations across systems.

Significance: The proposed virtual-to-clinical strategy is scalable and generalizable, offering a practical path to standardized CXR appearance and reliable downstream detection across institutions.