Rotator cuff injury (RCI) is a common cause of shoulder dysfunction, often accompanied by pain, restricted mobility, and reduced quality of life. Conventional imaging-based diagnostic approaches are effective for structural assessment but are often costly, time-consuming, and difficult to apply in large-scale screening or repeated follow-up. In this study, we propose the RCI recognition network (RCIRNet), a lightweight deep learning model for automated RCI recognition using upper-limb kinematic data collected from wearable inertial measurement units (IMUs). RCIRNet extracts temporal motion features with one-dimensional convolutional layers and introduces a channel-sequence cross-attention (CSCA) module to emphasize informative sensor channels and discriminative movement phases. Multi-segment acceleration signals from the upper thoracic spine, upper arm, and forearm during shoulder abduction are used to distinguish RCI from non-RCI conditions, including adhesive capsulitis and shoulders without dysfunction. Model performance is evaluated on an independent test set using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC). RCIRNet achieves an accuracy of 88.8%, an F1-score of 88.9%, and an AUC of 0.932, outperforming traditional machine-learning models as well as BiLSTM and Transformer Encoder baselines. These results suggest that IMU-based deep learning with channel-sequence attention provides a promising auxiliary tool for low-cost and non-invasive RCI screening.
Background and Objective: Fine-grained classification of instructed respiratory conditions from non-invasive electroencephalography (EEG) remains challenging because of low signal-to-noise ratio and substantial inter-participant variability. This study proposes a Deep Spatio-Respiratory Transformer (DSRT) for EEG-based classification of instructed respiratory conditions. Methods: DSRT integrates differential entropy (DE) and normalized power-ratio (NPR) features, squeeze-and-excitation-based feature recalibration, a four-layer feature-token Transformer encoder, and auxiliary nasal/oral pathway supervision. The final task comprised eight respiratory-condition classes. The framework was evaluated using 64-channel EEG from 13 unique participants contributing 21 recording sessions. A fixed-count window-selection strategy was used to equalize the number of retained EEG windows across classes, and model evaluation was performed using participant-disjoint outer folds. Primary architecture comparisons and feature-representation analyses were conducted using participant-level performance summaries. Results: At the participant level, DSRT achieved an Accuracy of 0.7429 ± 0.0492. Compared with EEG Conformer, the strongest matched-input comparison architecture, DSRT improved Accuracy and Macro-F1 by 3.78 and 3.79 percentage points, respectively. Architectural ablation showed the largest performance reduction after removal of the Transformer encoder, while participant-level feature-sensitivity analysis showed that the combined DE–NPR representation outperformed either feature family alone. DSRT contained 1.297 million parameters and achieved a median window-to-prediction latency of 3.825 ms. Duration-matched analysis further identified spatially heterogeneous EEG differences associated with instructed nasal and oral breathing. Conclusion: DSRT provides a participant-disjoint framework for fine-grained classification of EEG patterns associated with instructed respiratory conditions while retaining millisecond-scale inference. The findings support the utility of hybrid spectral representations and feature-token dependency modeling.
Brain-computer interfaces (BCIs) achieved by motor imagery (MI) have been widely accepted as an effective tool in rehabilitative applications. Functional near-infrared spectroscopy (fNIRS) and electroencephalography (EEG) are two common implementation forms of BCIs, capable of monitoring neuronal potential changes and hemodynamic variations during different tasks. MI and action observation are often employed in therapeutic settings for patients with hemiplegia or other conditions that prevent actual movements. Decoding physiological signals during the imagination process plays a role in facilitating rehabilitation. However, the role of guided motor imagery in the imagination process has not been well studied. In this study, we compared the activation levels of oxygenated hemoglobin, the changes of neural activities’ power in terms of time and frequency, and the differences and similarities of neural activations in the cortex area across seven different motor tasks under four conditions: motor execution (ME), MI, picture-guided motor imagery (PGMI) and video-guided motor imagery (VGMI). These findings demonstrate that although guided motor imagery (GMI) does not ease imagination tasks according to participant feedback, GMI reliably recruits mirror neurons during imagination when analyzing fNIRS and EEG signals. The results highlight that GMI has the potential to enrich BCI-based rehabilitation protocols for effectiveness enhancement.
Objective: Neuromuscular electrical stimulation (NMES) is a promising modality for restoring sensory feedback in upper-limb prosthetics. However, the dynamic cortical mechanisms underlying the perception of NMES-evoked muscle contraction, particularly under conditions of sensory loss (e.g., amputation), remain poorly understood. This study investigates the time-varying effective connectivity of the sensorimotor cortex in response to NMES using the ischemic nerve block (INB) technology to model acute deafferentation. Methods: 64-channel EEG data were recorded from 11 able-bodied participants under a 2×2 factorial design involving INB states (before vs. block) and stimulation intensities (above vs. below motor threshold). The adaptive directed transfer function (ADTF) combined with graph theoretical analysis was employed to quantify dynamic causal interactions in the Beta band (15-30 Hz). Results: The NMES-evoked cortical response progressed through three stages with monotonically increasing network segregation and integration (PFDR ≤ 0.01 for all stage-wise comparisons). During the ischemic nerve block (INB), the clustering coefficient (CC) and global efficiency (Ge) decreased under above-threshold stimulation (cluster-level PFDR = 0.019 and PFDR = 0.013, respectively), whereas below-threshold stimulation elicited a transient CC increase (150–525 ms, PFDR = 0.019) followed by late Ge reductions. The differentiation of the network response between above and below motor-threshold stimulation was reconfigured rather than abolished during deafferentation. At the regional level, deafferentation attenuated contralateral somatosensory and ipsilateral motor driving, whereas ipsilateral somatosensory outflow was relatively enhanced. Conclusion: NMES-evoked muscle contractions modulate the dynamics of sensorimotor networks by acute deafferentation. Acute deafferentation attenuates network integration and increases the relative contribution of the ipsilateral hemisphere. Significance: These electrophysiological markers may inform the design and evaluation of closed-loop sensory feedback for neuroprostheses.
In upper limb neuro-rehabilitation, decoding muscle synergies as neural movement primitives is essential for physiologically grounded assist-as-needed control. However, reconstructing these intention-rich patterns from non-invasive kinematics remains challenging due to the ill-posed nature of the inverse mapping problem and substantial inter-subject heterogeneity. To address these limitations, this paper proposes a ResMamba-based cross-modal distillation framework for robust muscle synergy reconstruction from planar single-joint upper-limb kinematics. Specifically, a hybrid temporal backbone integrating residual convolutions and Mamba selective state space models is constructed, enabling efficient sequence-level dependency modeling within the context window of non-stationary movement dynamics with linear computational complexity. To further regularize the ill-posed inverse mapping, an asymmetric teacher-student distillation strategy is designed, which transfers sEMG median frequency neuromodulation priors to a kinematics-only student network via soft labels during training. Furthermore, a two-stage transfer learning mechanism with 15% few-shot fine-tuning is developed to effectively overcome subject-specific heterogeneity with minimal calibration cost. Preliminary validation on planar single-joint tasks involving 10 healthy subjects and 3 stroke survivors with mild upper extremity impairment (FMA-UE: 56–59) shows that the proposed method achieves average prediction correlation coefficients of 0.857 (wrist) and 0.808 (elbow) in healthy scenarios. Furthermore, it achieves robust prediction correlation coefficients of 0.795 and 0.741 under mild pathological conditions, corresponding to an improvement of more than 20% over both the conventional single-stream CNN and Transformer baselines. This work provides a promising kinematics-driven intention decoding framework for single-joint rehabilitation in patients with mild motor impairment.
