IEEE Transactions on Neural Systems and Rehabilitation Engineering最新文献

Dysphagia is a common complication among stroke patients, significantly increasing the risk of aspiration pneumonia, malnutrition, and mortality. Traditional diagnostic techniques, such as bedside screening and videofluoroscopic swallowing studies, are limited by accessibility, reliability, and invasiveness. To address the challenges of limited data and complex multimodal signals, we propose a large language model (LLM)-based framework for dysphagia screening. This framework integrates multimodal physiological signals-including laryngeal vibration, nasal airflow, and swallowing sound-and leverages the powerful reasoning capabilities of LLMs for analysis. A medically-informed prompt template is designed to incorporate individual attributes, key biosignal features, and task instructions, effectively guiding the LLM to focus on dysphagia-related patterns. A total of 217 participants were recruited in this study, including 109 post-stroke patients with dysphagia and 108 healthy individuals, generating 1,391 dysphagic and 1,273 healthy control samples. Evaluation demonstrates that the proposed method achieves a classification accuracy of 96.3%, significantly outperforming baseline models. Notably, the model maintains robust performance in few-shot learning settings, indicating strong generalization capabilities. The proposed LLM-based framework offers a promising solution to early-stage clinical dysphagia screening by effectively integrating multimodal biosignals and leveraging prompt-driven reasoning, with extensive applicability in clinical practice.

Behavioral and psychological symptoms of dementia pose challenges to the safety and well-being of individuals in residential care. The integration of video surveillance in common areas of these settings presents a valuable opportunity for developing automated deep learning methods capable of identifying such behavior of risk. By issuing real-time alerts, these methods can support timely staff intervention and reduce the likelihood of incidents escalating. However, a persistent limitation is the considerable drop in performance when these methods are deployed in environments unseen during training. To address this issue, we propose an unsupervised scene-invariant fusion-based deep learning network. It combines language model-based captioning and scoring with video anomaly detection scoring to improve the generalization performance for unseen camera scenes. The video anomaly detection scoring uses a depth-weighted spatio-temporal autoencoder to reduce false positives, and the caption-based scoring uses a large language model to generate anomaly scores from captions of video frames. The study uses video data collected from nine individuals with dementia, recorded via three distinct hallway-mounted cameras in a dementia unit. The performance was investigated in both the same camera and cross-camera settings, where the proposed method performed consistently better than the existing methods. The proposed approach obtained the best area under receiver operating characteristic curve performance of 0.855, 0.84 and 0.805 for the three cameras. This work motivates further research to develop cross-camera behavior of risk detection systems for people with dementia in care environments.

Supernumerary robotic fingers (SRFs) are wearable assistive devices, which are increasingly incorporated into robotic rehabilitation programs aimed at restoring upper-limb function and promoting task-specific compensation. Despite growing evidence of SRF efficacy in improving motor performance, limited attention has been given to physiological adaptation and autonomic nervous system (ANS) integration during SRF use. This study investigated phase coherence (PC) and amplitude-weighted phase coherence (AWPC) of RR intervals derived from photoplethysmogram (PPG) as noninvasive biomarkers for autonomic nervous system adaptation during SRF-assisted activities of daily living. Thirty healthy participants completed a baseline (no SRF), pre-training SRF application and post-training SRF use, including rest periods protocol. Drinking water, driving and shape sorting were the functional activities of daily living (ADLs) that had to be completed. The results for PC and AWPC in the low (0.04-0.15) and high (0.15-0.4) frequency bands indicated an overall significant reduction in stress associated with SRF use (p <0.05). During the shape sorting task, post-training AWPC was significantly higher than in the pre-training phase (p = 0.037), and PC also increased significantly (p = 0.044), indicating enhanced vagal modulation. Driving task AWPC improved in the high-frequency band increasing from $0.68~pm ~0.12$ (no SRF) to $0.74~pm ~0.10$ (pre-training SRF) and $0.79~pm ~0.09$ (post-training SRF), while PC increased from $0.54~pm ~0.11$ to $0.62~pm ~0.08$ after training demonstrating significant task, phase, and frequency-specific alterations in autonomic coherence. This work provides an innovative perspective on physiological embodiment and how robotic compensation/augmentation improve both motor performance and physiological regulation. PD analysis indicated central autonomic adaptation. The current findings support the integration of coherence-based autonomic measures into assistive device evaluation frameworks to optimize training protocols and personalize robotic rehabilitation strategies.

Hemifacial spasm (HFS) is a refractory neuromuscular disorder that primarily affects facial movement. Electromyography (EMG) is one common-used clinical evaluation technique for HFS, which can help support the effective treatment. Nonetheless, due to its invasiveness and low-resolution issues, the application of EMG is limited. This study aimed to evaluate the facial muscle activities in HFS, by employing a whole-face surface EMG (sEMG) technique. Twenty patients with HFS were recruited in this study. Whole-face sEMG electrodes were used to record the muscle activities of the frontalis, orbicularis oculi, orbicularis oris, mentalis, and midfacial muscles during different facial movements. The root mean square (RMS) and median frequency (MDF) features from sEMG signals were computed for analysis. Especially, the sEMG topographic map was applied to analyze the coordinated activities of midfacial muscles, by using stretchable sEMG electrode arrays. Results demonstrated that the RMS of the affected side's frontalis was significantly higher than the unaffected side ( ${P} lt 0.001$ ), while no significant difference was observed in MDF ( ${P} gt 0.05$ ). In the RMS topographic map, the center of gravity in the horizontal direction (CoGx) shifted significantly toward the midline on the affected side during the teeth-showing task ( ${P} lt 0.05$ ) and the entropy feature on the affected side was significantly lower than that on the unaffected side during the cheek-showing task ( ${P} lt 0.05$ ). These findings indicate the compensatory mechanisms in the facial nerve distribution areas of HFS and provide an effective evaluation tool for the objective quantification of HFS severity and abnormal co-activation patterns.

Lower-limb rehabilitation benefits from an input channel that conveys intent cleanly while actuation remains predictable and safety-bounded. We present a clinic-friendly, binocular gaze-driven paradigm that maps quadrant fixations to discrete commands for a two-degree-of-freedom ankle robot (dorsiflexion/plantarflexion and internal/external axial rotation), while inversion/eversion can be left compliant or mechanically constrained as needed. Pupil centers from a near-infrared tracker are mapped to a unit-normalized screen plane using low-order regression with ArUco-guided homography and rapid affine correction. A conservative dwell/occupancy rule triggers jerk-limited trajectories executed under cascaded position-velocity-current control with software rate/torque limits and watchdog supervision. In 20 healthy adults (1,600 trials), selection accuracy reached 99.94% with a 157.6 ms median end-to-end delay (gaze onset to motor onset). A head-tremor stress test produced no wrong-quadrant decisions and withheld decisions at the highest severity when the occupancy criterion was not met. Under passive drives, tracking was sub-degree (RMSE $le 0.224^{circ } $ ) with smooth profiles and torques within software limits. Human factors outcomes were favorable, including a pilot post-stroke cohort, with high usability, low workload, and minimal visual fatigue ( $Delta $ VAS 0.14/0.21). These results support gaze as a practical, hands-free primary control channel for seated ankle training in clinical workflows.

Robotic systems assess joint dynamics objectively by perturbing the limb and estimating properties such as impedance. Position perturbations constrain the limb to a target trajectory, reducing variability in task execution but obstructing voluntary motion. Force perturbations allow voluntary movement but elicit orientation-dependent responses, increasing the number of trials needed for accurate estimates. To overcome these limitations, we combined the flexibility of admittance control with the repeatability of position perturbations. A minimum-jerk trajectory ensures smooth transitions. The experiment with six healthy participants was performed to demonstrate the reliability, accuracy and smoothness of applying such perturbations during voluntary movement. Reliability was the proportion of perturbations that reached the target velocity within one millisecond of the acceleration time window. Accuracy was measured as the RMSE between the target and measured velocity during the constant velocity. Smoothness was assessed as perceivability: the fraction of trials in which participants correctly detected a perturbation. The controller allows continuous voluntary movement, switching only during perturbations to impose a precise, specified perturbation. All perturbations reached the target velocity within one millisecond of the acceleration time window; thus, the method is reliable. Under the most demanding condition-an increase to 200 deg/s in 0.01 s-the RMSE between target and measured velocity was 1.1 deg/s (0.55%), indicating a high accuracy. Specially designed perturbations had a perceivability accuracy of 22.1%, indicating smooth transitions between control modes. Together, these results indicate a promising approach for assessing joint dynamics during voluntary elbow movement, enabling assessment during activities of daily living.

Neuromuscular electrical stimulation (NMES) is widely employed for restoring upper limb motor function post-stroke. However, precise control of fine hand movements remains a significant challenge due to the poorly understood time-varying dynamics governing the translation of electrical stimulation into muscle force and joint motion. This study investigated the dynamic modulation of muscle contraction and finger kinematics under varying NMES durations (0.4s, 0.6s, 1s) in 11 healthy subjects. To decouple evoked muscle contractions from joint motion feedback, Ischemic Nerve Block (INB) was employed, after which NMES (0.4s and 1s) was re-administered. Kinematic metrics (angular acceleration and jerk) of evoked finger movements and wavelet spectral features of mechanomyography (MMG) signals were extracted to quantify the time-varying dynamics of NMES-evoked finger movements and/or contraction force dynamics. Results showed that while angular acceleration increased logarithmically with stimulation duration, movement smoothness-quantified by jerk-exhibited a quadratic polynomial decay. Correspondingly, MMG wavelet spectral features (3-20 Hz) exhibited a quadratic polynomial decay as stimulation duration increased, reflecting distinct phases of motor unit recruitment and saturation. A quadratic polynomial correlation (R ${}^{{2}} =0.2574$ ) between jerk and MMG features confirmed that the smoothness of finger motion is directly dictated by the intrinsic mechanical oscillation of muscle fibers. Importantly, these patterns persisted after INB, demonstrating that NMES modulates fine motor output primarily through intrinsic muscle force dynamics rather than sensory feedback loops. These findings provide a physiological basis for optimizing stimulation parameters to achieve smooth, force-regulated control of paralyzed hands.

Accurate assessment of gross motor function in children with cerebral palsy (CP) is essential for clinical decision-making, yet current practice is limited by data scarcity, severe class imbalance, and patient heterogeneity. Recent skeleton-based deep learning approaches, such as spatio-temporal graph convolutional networks (STGCN), enable automatic GMFCS prediction from monocular video but are limited in generalizability and fairness. In this work, we propose a unified generative-diagnostic pipeline that integrates a Conditional Skeleton Diffusion Model (CSDM) with a Biomechanically-Aware Spatio-Temporal Graph Convolutional Network (BA-STGCN). The CSDM generates clinically plausible 2D skeleton gait sequences conditioned on Gross Motor Function Classification System (GMFCS) level, Gait Deviation Index (GDI), and anthropometric covariates, guided by an anatomically structured covariance model to preserve biomechanical fidelity and clinical distributions. These synthetic sequences, combined with real patient data, are used to train the BA-STGCN, which incorporates a symmetry-based loss and a multi-task head for joint GMFCS classification and continuous GDI regression. Extensive evaluation on a pediatric clinical gait dataset demonstrates that our approach achieves 85.7% GMFCS classification accuracy with balanced precision and recall, reduces mean absolute error in GDI prediction to 4.6, and markedly improves recognition of severe phenotypes. These findings highlight that conditional skeleton diffusion, coupled with biomechanically informed graph learning, provides a scalable, interpretable, and privacy-preserving pathway for automated clinical gait assessment in CP.

Electromyography (EMG) is essential in medical and rehabilitation fields for assessing neuromuscular functions. However, mainstream traditional surface EMG (sEMG) is susceptible to electrode displacement or noise interference during walking, leading to lower signal quality and limited long-term stability, which constrains its broader application. To overcome these limitations, we propose a novel intravascular electromyography (iEMG) acquisition method. By a minimally invasive surgery, a self-expanding stent sensor with microelectrode and lead was implanted into the femoral vein adjacent to the tibialis anterior muscle to record deep muscle activity. In this study, sEMG and iEMG were simultaneously acquired from sheep hindlimb muscle in standing state and walking state for comparative analysis of their electrophysiological properties. Level walking served as the dynamic locomotor task in this work. Totally, this acquisition experiment lasted for three days. In standing state, both EMG exhibit a high correlation (Spearman's $rho ={0}.{9018}$ , ${p}lt {0}.{001}$ ). In walking state, iEMG demonstrates a 10.04% higher signal-to-noise ratio compared to sEMG. Additionally, iEMG shows a 33.82% lower coefficient of variation in power spectral density than sEMG, indicating a 1.51-fold improvement in signal stability. These results demonstrate that our iEMG acquisition method enables high-quality and robust recordings for long-term monitoring during walking, which provides a reliable foundation for clinical rehabilitation applications requiring precise, long-term muscle activity tracking.

Aphasia is a common condition following brain injury, traditionally assessed and treated by speech therapists through manual evaluations and conventional language rehabilitation. However, these methods are time-consuming, reliant on professionals, and subject to subjective biases. This study aims to develop a virtual speech therapy room with an automated assessment model to assist clinicians in evaluation. It provides immersive virtual reality (VR) language training modules, combining analysis of physiological data to achieve the goal of smart healthcare. Twenty individuals with aphasia (IWA) and ten healthy participants were involved, with aphasia subjects randomly assigned to the experimental and control group A, and healthy participants forming control group B. Clinical scales, VR tasks, and neurobehavioral data were measured as needed. Statistical analysis confirmed that using virtual reality can enhance the effectiveness of aphasia treatment interventions. Utilizing virtual reality and behavioral sensing technology, significant differences were observed in the left frontal and occipital regions between IWA and healthy participants, aligning with clinical observations of impaired language and visual processing areas. The assessment model, established through these data, achieved an average classification accuracy of 97% in distinguishing between individuals with aphasia and healthy participants using multimodal fusion with repeated cross-validation, indicating its potential as an auxiliary tool for physician assessment and treatment.