IEEE Transactions on Biomedical Engineering最新文献

Background: Acquiring human- like motor skills in embodied musculoskeletal models is challenging due to the high dimensionality and redundancy of muscle actuators.

Methods: Inspired by human motor control coordination patterns, we introduce a synergy-guided reinforcement learning framework integrating physiological priors derived from muscle synergies into the control policy of an embodied musculoskeletal model. It leverages coordinated muscle activation patterns to guide learning, generating muscle excitation signals via a synergy-guided control component and a residual control component. To evaluate the proposed method, four badminton stroke skills are selected as benchmark tasks (forehand/backhand, inward/outward net slices).

Results: The experimental results demonstrate that our method achieves an average root mean square error of under 0.015 radians across all stroke types, demonstrating its ability to accurately learn expert motion. Furthermore, it outperforms the baseline proximal policy optimization (PPO) model in terms of trajectory accuracy, energy efficiency, and training convergence speed, particularly excelling in energy efficiency with up to a 14.9% reduction in energy consumption. The forehand high serve is also tested to validate the method's effectiveness in learning longer and larger-ranges movements, showing the same advantages. Moreover, the muscle synergies learned by the model exhibit moderate resemblance to human synergies, indicating potential interpretability and biological plausibility.

Conclusion: This work highlights that integrating neurophysiological priors into reinforcement learning provides a promising pathway for efficient, interpretable, and human- like motor control.

Significance: The approach holds promise for advancing motor skill assessment, human-machine interfaces, and rehabilitation technologies by enabling more efficient and human- like motion skill learning.

Objectives: Managing acute heart failure (AHF) involves the dual challenges of stabilizing hemodynamics while minimizing myocardial oxygen consumption. This is a clinically significant and technically challenging issue due to complex cardiovascular interactions and trade-offs between therapeutic goals. We aim to develop a discrete hemodynamic control framework for autonomous drug therapy that enhances both control versatility and clinical potential.

Methods: A comprehensive multi-input, multi-output model was derived to capture causal relationships linking drug infusion, cardiovascular parameters, and hemodynamic responses. Based on this model, we designed an optimal control framework capable of regulating arbitrary hemodynamic targets (e.g., arterial pressure, cardiac output, left atrial pressure, right atrial pressure), minimizing myocardial oxygen consumption, accommodating multiple guideline-concordant drug classes, and enforcing clinical constraints such as dosage limits and contraindications. The discrete design enables compatibility with standard hemodynamic measurements in ICU/CCU settings. Control performance was validated using a cardiovascular simulator.

Results: The system achieved multi-dimensional regulation of hemodynamics and reduced myocardial oxygen consumption across a range of AHF scenarios. It also identified pharmacologically untreatable cases by evaluating the reachable hemodynamic range. Performance remained stable despite inter-patient variability in drug sensitivities, circulatory properties, and baroreflex activity.

Conclusions: The proposed system may function as a physician-assistive tool and a foundation for autonomous drug therapy in AHF.

Significances: By integrating physiologically grounded modeling with practical clinical constraints, the system offers a novel and scalable approach to optimize drug therapy in AHF. It has the potential to improve treatment precision, reduce clinician burden, and advance next-generation critical care.

The electrocardiogram (ECG) is a vital diagnostic tool used to monitor and diagnose a wide range of cardiac conditions. However, ECG signals can exhibit significant variability across different patient populations, recording devices, and environmental conditions, creating challenges in developing universally robust and accurate classification models. This research addresses these challenges by exploring and advancing domain adaptation techniques to enhance the robustness and generalizability of ECG classification models. By leveraging unsupervised domain adaptation (UDA), we aim to mitigate the performance degradation that typically occurs when models trained on one dataset are applied to another, thereby improving diagnostic accuracy and reliability across diverse clinical settings. We introduce ECG-Adapt, an integrated approach that aligns features both within classes and across domains. Unlike existing methods that rely on clustering as a preprocessing step, ECG-Adapt does not require clustering, simplifying the workflow. It further incorporates weakly supervised learning to prevent overfitting of the discriminator to pseudo-labels generated by the classifier, enhancing robustness and generalizability. Applying our novel unsupervised domain adaptation framework led to substantial performance gains. For instance, ECG-Adapt improved the average $F_{1}$-score by 8% on single-lead problems and 7% on 12-lead problems. By leveraging ECG-Adapt, performance degradation when applying models across datasets can be mitigated, enhancing diagnostic accuracy and reliability in diverse clinical settings and demonstrating strong potential for real-world deployment.

Objective: Upper-limb electrocardiogram (ECG) acquired from a single-arm wearable device offers a practical approach for dynamic cardiac monitoring. This study addresses whether single-channel Arm-ECG, characterized by higher noise and distinct morphology compared to standard 12-lead ECG, can reliably detect arrhythmias.

Methods: We propose CLMF-Net, a contrastive multitask framework featuring multiscale convolutional layers to capture temporal-morphological patterns. A fine-grained reconstruction branch preserves subtle clinical features, including P waves and ST segments, while a contrastive module ensures robustness against signal quality variations. The complete framework is trained end-to-end using a unified loss function that balances classification, reconstruction, and consistency objectives.

Results: Validation with 132 elderly participants undergoing six-minute walk tests confirmed the feasibility of dynamic Arm-ECG acquisition. CLMF-Net, trained for arrhythmia identification from single-channel ECG, achieved 95.19% accuracy on the Arm-ECG dataset. It effectively detects arrhythmias, despite challenges with noise and morphology. Generalizability was evidenced by 95.55% and 84.23% accuracy on the Chapman and China Physiological Signal Challenge 2018 (CPSC2018) datasets, respectively, demonstrating the model's ability to classify multiple arrhythmia types across diverse datasets.

Conclusion: Beyond precise identification of arrhythmias, deep feature and morphology analyses showed that the learned representations do not always align with clinically emphasized intervals. This discrepancy highlights both the promise of leveraging Arm-ECG signals and the caution required when translating model-derived interpretations into clinical workflows.

Significance: This study validates the clinical utility of single-arm ECG monitoring, demonstrating feature reliability for real-world cardiac health assessment and supporting the translational potential through consistency with standard recordings.

Objective: Inter-session and inter-subject variability in electroencephalography (EEG) signals, resulting from individual differences and environmental factors, poses a major challenge for neural decoding in brain-computer interface (BCI) applications.

Methods: To address this issue, we propose RUNet, a zero-calibration motor imagery EEG decoding framework based on Riemannian manifold learning and unsupervised representation learning. RUNet incorporates a multi-scale spatiotemporal convolutional module that jointly captures local global spatial and multi-resolution temporal dynamics features. To enhance the robustness of EEG features against non stationarity, a polysynergistic covariance optimization module is employed, which strengthens the covariance matrix representation through multiple regularizations and adaptive fusion. In addition, RUNet integrates the Riemannian Affine Log Mapping layer, based on Affine-Invariant Transformation and Log-Euclidean Mapping, in an end-to-end manner to mitigate cross-domain covariance drift and enhance domain-invariant feature learning. A transfer learning framework is further integrated into RUNet: during pre-training, an unsupervised contrastive loss is applied to resting-state EEG data to learn domain-invariant spatiotemporal features; during retraining, task-specific data are used to enhance discriminability and feature disentanglement.

Conclusion: Experimental results on the BCI Competition IV 2a, 2b datasets and a self-collected laboratory dataset show that RUNet achieves average cross-session accuracies of 87.19%, 88.03% and 85.45%, and cross-subject accuracies of 68.09%, 78.29% and 87.25%, respectively. On the PhysioNet dataset, a cross-subject accuracy of 78.14% is achieved. These results demonstrate the effectiveness of RUNet's unified pipeline and its robust cross-domain generalization.

Segment Anything Model (SAM) has gained significant attention because of its ability to segment a variety of objects in images upon providing a prompt. Recently developed SAM 2 has extended this ability to video segmentation, and by substituting the third spatial dimension in 3D images for the time dimension in videos, it opens an opportunity to apply SAM 2 to 3D medical images. In this paper, we extensively evaluate SAM 2's ability to segment both 2D and 3D medical images using 80 prompt strategies across 21 medical imaging datasets, including 2D modalities (X-ray and ultrasound), 3D modalities (magnetic resonance imaging, computed tomography, and positron emission tomography), and surgical videos. We find that in the 2D setting, SAM 2 performs similarly to SAM, while in the 3D setting we observe that: (1) selecting the first mask is more effective than choosing the one with the highest confidence, (2) prompting the slice with the largest object appears is the most cost-effective strategy when only one slice is prompted, (3) box prompts result in higher performance than point prompts at a slightly higher annotation cost, (4) bidirectional propagation outperforms front-to-end propagation, (5) interactive annotation is rarely effective, (6) SAM 2, without fine-tuning, achieves 3D IoU from 0.32 with a single point prompt to 0.51 with a ground truth mask on one slice, and exceeds 0.8 on certain datasets when using box or ground-truth prompts, a level that begins to approach clinical usefulness. These findings demonstrate that SAM 2's ability to segment 3D medical images can be improved with our proposed strategies over the default ones, providing practical guidance for using SAM 2 for prompt-based 3D medical image segmentation.

Quantitative viscoelasticity imaging via shear wave elastography (SWE) remains challenging due to complex wave physics and limitations of conventional reconstruction methods. To address this, we present SW-VEI-Net, a physics-informed neural network (PINN) that simultaneously reconstructs the shear elastic modulus and viscous modulus by integrating viscoelastic wave equations into a dual-network architecture. The framework employs a dual-loss function to balance data fidelity and physics-based regularization, significantly reducing reliance on empirical data while improving interpretability. Extensive validation on tissue-mimicking phantoms, rat liver fibrosis model, and clinical cases demonstrates that SW-VEI-Net outperforms state-of-the-art SWE methods. Compared to SWENet (a PINN-based method using a linear elastic model), SW-VEI-Net not only enables simultaneous assessment of shear elastic and viscous moduli, but also achieves higher accuracy in shear elastic modulus reconstruction. Furthermore, when benchmarked against the dispersion fitting (DF) method (based on a viscoelastic model), SW-VEI-Net produces comparable viscoelastic parameter maps while exhibiting enhanced robustness and consistency. For liver fibrosis staging, SW-VEI-Net achieves AUC values of 0.85 ($geq$F2) and 0.91 ($=$F4) based on elastic modulus classification, surpassing both SWENet (0.84, 0.85) and DF (0.78, 0.88). Additional validation in healthy volunteers shows strong agreement with a commercial ultrasound system. By synergizing deep learning with fundamental wave physics, this study represents a significant advancement in SWE, offering substantial clinical potential for early detection of hepatic fibrosis and malignant lesions through precise viscoelastic biomarker mapping.

Pulse wave velocity (PWV) is an indirect measure of vessel stiffness, that has the potential to serve as a meaningful parameter for risk stratification of vascular diseases, such as abdominal aortic aneurysms (AAAs). However, assessing the PWV and pulse wave patterns in the complete abdominal aorta using ultrasound-based pulse wave imaging (PWI) is challenging due to the limited field of view (FOV) and contrast of a single ultrasound (US) probe. Hence, an approach is required that can capture distension of aortas with different levels of stiffness accurately in a large FOV. Therefore, we propose PWI based on dual probe, bistatic US. Single and dual probe ultrasound simulations were performed using finite element models of pressure waves propagating in aortas with different stiffness levels. Next, the approach was tested on an aorta and AAA mimicking phantom in a mock circulation setup. The simulation results show that the FOV, image quality, and PWV-estimation accuracy improve when using the dual probe approach (accuracy range: 94.9 - 99.8 $%$; R$^{2}$ range: 0.92 - 0.98) compared to conventional US (accuracy range: 12.6 - 93.9 $%$; R$^{2}$ range: 0.52 - 0.91). The approach was successfully expanded to the phantom study, which demonstrated expected wave patterns within a larger FOV. With dual probe PWI of the non-dilated phantom, the R$^{2}$-value improves (monostatic: 0.95; bistatic: 0.96) compared to use of single probe PWI (0.85). The proposed method shows to be promising for PWV-estimations in less compliant vessels with high wave speeds.

Multimodal signal fusion is a cornerstone of biomedical engineering and intelligent sensing, enabling holistic analysis of heterogeneous sources such as electroencephalography (EEG), peripheral signals, speech, and imaging data. However, integrating diverse modalities in a computationally efficient and biologically plausible manner remains a significant challenge. Transformer-based fusion architectures rely on global cross-attention to integrate multimodal information but incur high computational costs. In contrast, STDP-driven fully connected layers adopt local learning rules, which restrict their ability to autonomously form efficient sparse topologies for complex multimodal tasks. To address these issues, we propose a novel end-to-end framework-the Multimodal Spiking Neural Network (MSNN)-featuring a fusion module grounded in the Generalized Distributive Law (GDL). This principled mechanism provides an efficient and interpretable means of integrating heterogeneous biomedical and sensory signals. The MSNN further incorporates structure-adaptive leaky integrate-and-fire (SALIF) neurons, enabling dynamic optimization of sparse connectivity to enhance fusion efficiency. The proposed MSNN is validated on a range of datasets, demonstrating strong versatility: it achieves binary classification accuracies of 92.29% (valence) and 91.08% (arousal) on the DEAP dataset for affective state decoding and 99.77% on the WESAD dataset for stress detection, while delivering state-of-the-art performance on standard pattern recognition tasks (MNIST & TIDIGITS: 99.01%) and event-driven neuromorphic datasets (MNIST-DVS & N-TIDIGITS: 99.98%). These results demonstrate that MSNN offers an effective and energy-efficient solution for multimodal sensor fusion in biomedical and intelligent sensing applications.

Ultrasound (US)-guided needle tracking is a critical procedure for various clinical diagnoses and treatment planning, highlighting the need for improved visualization methods to enhance accuracy. While deep learning (DL) techniques have been employed to boost needle visibility in US images, they often rely heavily on manual annotations or simulated datasets, which can introduce biases and limit real-world applicability. Photoacoustic (PA) imaging, known for its high contrast capabilities, offers a promising solution by providing superior needle visualization compared to conventional US images. In this work, we present FocFormer-UNet, a DL network that leverages PA images of the needle as ground truth for training, eliminating the need for manual annotations. This approach significantly improves needle localization accuracy in US images, reducing the reliance on time-consuming manual labeling. FocFormer-UNet achieves excellent needle localization accuracy, demonstrated by a modified Hausdorff distance of 1.43 1.23 and a targeting error of 1.22 1.14 on human clinical dataset, indicating minimal deviation from actual needle positions. Our method offers robust needle tracking across diverse US systems, improving the precision and reliability of US-guided needle insertion procedures. It holds great promise for advancing AI-driven clinical support tools in medical imaging. The following is the source code: https://github.com/DeeplearningBILAB/FocFormer-UNet. Open Science Framework (OSF) provides datasets and checkpoints at: https://osf.io/yxt9v/.