IEEE Transactions on Neural Systems and Rehabilitation Engineering

There are many alternative methods to joystick control for control of Electric Powered Wheelchairs for users with neuromuscular disabilities, such as muscular dystrophy, and spinal cord injuries, such as tetraplegia. However, these methods- which include the sip-and-puff method, head and neck movement, blinking, or tongue movement- hinder social interaction, and are therefore detrimental to user independence. In recent years, research has explored the use of Electromyography (EMG) signals from alternative muscles to control a powered wheelchair, consequently increasing the quality of life of these users. The Auricular Muscles (AM) may be suitable, as they are controlled separately from the facial nerve and are vestigial in humans, making them advantageous for powered wheelchair control for users with tetraplegia. Additionally, they are located around the ear, adding a level of cosmesis when designing wearable sensors and prosthesis. This paper extracts and implements two control strategies from current literature and, for the first time, compares them directly, demonstrating viable implementation approaches for an online EMG-based powered-wheelchair control system. A Support Vector Machine (SVM) was developed and various window lengths were compared, with the most accuracy and real-time effectiveness found at 300ms. A study with three participants demonstrates the feasibility of these methods of control as well as experimental results to guide the potential AM use.

Instrumented gait assessment in pediatric populations is often constrained by the complexity and lack of portability of traditional motion capture systems. In this article, we propose a deep learning approach utilizing a one-dimensional (1D) U-Net architecture to accurately estimate ankle and knee joint angles in the sagittal plane from surface electromyography (sEMG) signals. We analyzed data from the tibialis anterior and medial gastrocnemius of 25 typically developing children (ages 4-16) to evaluate the models performance and the influence of age-related gait maturation. The proposed 1D U-Net achieved high predictive accuracy for the ankle joint (RMSE: 3.6{degrees}) and the knee joint (RMSE: 4.1{circ}). Experimental results demonstrated that incorporating the toe-off event as a temporal marker significantly enhanced prediction stability during transitional gait phases. Furthermore, Statistical Parametric Mapping (SPM) was employed to identify systematic errors, which were primarily localized during initial contact and pre-swing but remained below clinically relevant thresholds. The findings reveal that prediction accuracy increases with age, reflecting more stable neuromotor patterns. This study demonstrates that a 1D U-Net can reliably decode lower-limb kinematics from sEMG alone, enabling the development of simplified, non-invasive, and portable pediatric gait assessment tools that can be integrated into the control strategies of assistive devices.

ObjectiveAccurate stride length measurement is essential for assessing functional mobility, yet gold-standard methods remain confined to laboratory settings. This study aimed to develop and validate a computationally efficient, interpretable linear model for predicting stride length using thigh- and shank-mounted inertial measurement units integrated into a wearable neuromodulation sleeve. MethodsData from the sleeve were collected from 29 healthy adults performing walking bouts at four self-selected speeds. Participants traversed a pressure-sensitive gait mat, providing gold standard labels. A linear regression model was developed from engineered features from the kinematics data streams and validated against a held-out test set (n = 6) using leaveone-participant-out cross-validation. ResultsThe final linear model utilized five predictors: participant height, shank range of motion (ROM), thigh ROM, and thigh swing duration metrics. It achieved high predictive accuracy with a mean absolute error (MAE) of 5.98 cm, a mean absolute percentage error (MAPE) of 4.53%, and an R2 of 0.89. The model significantly outperformed naive baseline models (p < 0.05) and performed similarly to more complex non-linear architectures, such as neural networks and random forests. Notably, 88.4% of strides were predicted within 10% of the ground truth. ConclusionA parsimonious linear model leveraging proximal limb kinematics provides accurate and biomechanically interpretable stride length estimation. Low computational demand makes it suitable for real-time, ondevice gait monitoring in wearable assistive technologies, facilitating clinical assessments in real-world environments.

Optimizing assistive wearable devices is crucial for their efficacy and user adoption, yet state-of-the-art methods like Human-in-the-Loop Optimization (HILO) and biomechanical modeling face limitations. HILO is time-consuming and often restricted to optimizing control parameters, while inverse dynamics assumes invariant kinematics, which is unreliable for adaptive human-device interaction. Predictive simulation offers a powerful alternative, enabling computational exploration of design spaces. However, existing approaches often lack systematic optimization frameworks and rigorous validation against experimental data. To address this, we developed a Design Optimization Platform that integrates predictive simulations within a two-level optimization structure for personalizing assistive device design. This paper primarily validates the platforms predictive simulations against a publicly available dataset of the passive Biarticular Thigh Exosuit (BATEX), assessing its reliability. Our findings show that the model can sufficiently predict the kinematics and major muscle activations, except for the pelvis tilt and some biarticular muscles. The key finding is that successful identification of personalized optimal BATEX stiffness parameters needs acceptable prediction of metabolic cost trends, not their precise values. Our analysis further reveals that the models accuracy in predicting Vasti muscle activation in the baseline condition is a significant indicator of its success in predicting metabolic cost trends. This demonstrates that accurate prediction of performance trends is more important for effective simulation-based design optimization than perfect biomechanical accuracy, advancing targeted and efficient assistive device development.

The kinematics of rhythmic, speed-modulated finger and grasp-like movements were analyzed using a reduced biomechanical model of the hand and a marker-based optical motion-capture system. Twenty-one healthy participants performed eight hand motor tasks involving metacarpophalangeal (MCP) joint flexion-extension (F-E) and carpometacarpal (CMC) thumb opposition-reposition (O-R) at two movement frequencies (0.50 and 0.75 Hz). Kinematic analysis quantified the range of movement (RoM), mean speed, and normalized total harmonic distortion (TDHN). Statistical analysis identified task type as the primary factor modulating all three metrics across digits, with large effect sizes [Formula]. Movement frequency significantly influenced mean speed [Formula] and moderately affected TDHN [Formula], while thumb RoM remained statistically unchanged across frequencies (p = 0.063). Participants consistently reproduced the intended sinusoidal trajectories, as indicated by low TDHN values (below 19%). The findings support the analysis of coordinated hand movements across various tasks under controlled time conditions. They also demonstrate that the simplified biomechanical model accurately captured both individual and co-ordinated finger movements. This provides a valuable reference for studies on motor control and for applications in rehabilitation and assistive technology.

Brain-Computer interfaces (BCIs) offer a link between neural signals and external computation, enabling control of devices for the purposes of restoring function to motor-affected individuals and enhancing capabilities of a wider set of populations. Electroencephalography (EEG) offers a high temporal resolution for dynamic and potential real-time feedback for non-invasive systems. However, its practical efficacy remains limited due to low spatial resolution and poor signal-to-noise ratio, leading to insufficient decoding accuracy and unintuitive control paradigms that hinder reliable user interaction. In this study, we present a framework for an online EEG foundation model by creating a custom foundation model through spectrogram reconstruction of compact temporal windows and online constraints during pretraining. We evaluate the performance of the model in a challenging control paradigm of single-arm, directional motor imagery with dynamic movements for guided and free movement cursor control tasks. Our foundation model approach achieved a final average accuracy of 51.3% during a goal-oriented guided control task. This represents a 15.8% increase over a conventional deep learning framework and a 26.3% increase above chance level, evaluated in a cohort of 11 human participants. During the free movement task, the foundation model invoked a higher rate of completion and lower completion times. Furthermore, the custom EEG foundation model demonstrated superior adaptability from same-session finetuning and indicated an enhanced capability to assist subject learning. These findings highlight the potential of EEG foundation models to support more robust and intuitive non-invasive BCI systems, providing a promising modelling framework for future BCI development.

Motor rehabilitation requiring sustained physical exercise faces poor adherence in neurological populations due to insufficient supervision and monotony. While virtual reality and musical biofeedback independently improve engagement and motivation, their comparative and combined impact on intensity control strategies during high-intensity interval training (HIIT) remains unexplored. Thirty healthy adults (16 males, 14 females; mean age 27.5 {+/-} 7.2 years) were sequentially assigned to three feedback modalities (n=10 each) during intensity-guided stationary cycling: visual-only (position-based), musical-only (speed-based), and combined audiovisual (position-based). Participants completed two 9-minute moderate-to-high intensity sessions (Set 1 and Set 2) maintaining pedaling speed within a target speed zone. Performance distinguished control strategy from effectiveness: stability via target zone exits, correction capacity via recovery time and sustained deviations, and overall effectiveness via time in zone. Heart rate (HR) assessed physiological intensity; usability and cognitive workload were evaluated via e-Rubric and NASA-TLX. Distinct regulation strategies emerged. Musical-only showed significantly lower stability (Set 1: 14.52 exits/min vs. 1.48 visual and 1.79 combined; corrected p < 0.0167) but superior correction (0.21s recovery vs. 2.48s and 1.06s; p < 0.0001) with minimal sustained deviations. Combined feedback achieved highest Set 2 effectiveness (98.13% vs. 95.17% time in zone; corrected p < 0.0167) but elevated physical demand (corrected p < 0.0167). HR variability was comparable (p = 0.85), confirming consistent cardiovascular workload despite differing strategies. Satisfaction was high, with slight preference for musical feedback; cognitive workload did not differ. Musical biofeedback promotes reactive control with frequent but rapidly corrected oscillations, maintaining physiological safety and engagement. Visual feedback ensures stable target adherence at the cost of compensatory physical effort. Combined modality offered no synergy, increasing demand without improving effectiveness. Findings reveal a trade-off between stability and correction agility, supporting tailored modality selection: musical feedback suits unsupervised rehabilitation prioritizing engagement, rapid error correction, and sustainable effort, while visual feedback suits supervised protocols requiring stable preventive control and precise adherence quantification. Author summaryMany people undergoing neurological rehabilitation struggle to maintain adherence to high-intensity exercise programs, particularly without direct supervision. While virtual reality and musical feedback have shown promise for improving engagement and motivation, we didnt know which type works best for controlling exercise intensity, or whether combining them would be better. We tested three feedback systems with 30 healthy adults performing stationary cycling: visual-only, musical-only, and both combined. We measured how well participants stayed within target speed and assessed their experience. Musical feedback prompted frequent but instant adjustments--a reactive strategy that was less physically demanding and most enjoyable. Visual feedback kept participants more precisely in the target zone but required significantly more effort. Surprisingly, combining both didnt improve performance and instead increased physical demand. Our results show that different feedback types suit different rehabilitation contexts. Musical feedback may be ideal for unsupervised home-based exercise because it keeps people engaged without requiring exhausting effort. Visual feedback works better when precise control is essential in supervised clinical settings, despite being more demanding. Combining both offers no advantage. These findings help clinicians choose the right feedback approach based on their specific rehabilitation goals.

Vertical ground reaction force (vGRF) profiles during walking typically exhibit a double-peaked structure with a mid-stance trough, yet the mechanical conditions governing this morphology remain incompletely defined. In this study, we examined how the balance between push-off and collision impulses during the step-to-step transition influences the temporal and structural characteristics of the vGRF trajectory. Empirical relationships describing push-off and collision work were used to compute transition impulses across walking speeds ranging from 0.8 to 1.4 m{middle dot}s{square}1. A normalized Impulse Balance Index (IBI) was defined to quantify the relative dominance of push-off and collision impulses. The temporal position of the mid-stance trough was quantified using a Trough Deficit Index (TDI) derived from quadratic fits of the vGRF trajectory. Across walking speeds, push-off and collision variations produced step-to-step active work performance imbalance. Push-off and collision became approximately balanced near 1.2 m{middle dot}s{square}1, corresponding to the mechanically preferred walking speed. Deviations from this balanced condition were associated with systematic shifts in trough timing: the trough occurred 1.83% and 1.56% earlier in stance at 0.8 and 1.0 m{middle dot}s{square}1, respectively, and 1.31% later at 1.4 m{middle dot}s{square}1 relative to the reference speed. TDI exhibited a strong inverse relationship with impulse balance (IBI), indicating that vGRF morphology is tightly coupled to the mechanical balance of the step transition. A simplified pendular model further demonstrated that active torque, representing work, during single support shifts the quadratic vertex of the force trajectory by approximately 48.6-51.1% of stance, consistent with the observed trough timing variations. These results show that vertical GRF morphology reflects the imbalance between push-off and collision provides a simple signal of step-to-step transition mechanics, that may be used for rehabilitation and exoskeleton modulation.

Neural feedback is important for healthy control of movement, and multiple neurological disorders (e.g., stroke, cerebral palsy, Parkinsons disease, incomplete spinal cord injury) can be described by how they impair healthy feedback or induce unhealthy feedback. Researchers have created numerous computational neuromusculoskeletal models controlled by simulated neural feedback mechanisms, but these models rarely represent actual human subjects and thus have not found practical application in treating patients with movement impairments. As a step toward designing patient-specific treatments for individuals with neurological disorders, this study used the Neuromusculoskeletal Modeling Pipeline to develop and evaluate a novel synergy-based feedforward (FF)+feedback (FB) model using a personalized, three-dimensional neuromusculoskeletal walking model of an actual human subject post-stroke. Experimental walking data collected from the subject were used to create the subjects personalized walking model. This model was used to calculate lower body muscle activations consistent with the subjects electromyographic, joint motion, and ground reaction data for 5 calibration walking cycles. Nominal FF synergy controls were calculated by averaging the muscle synergies that closely reconstructed the 5 cycles of muscle activations and associated joint moments simultaneously. These nominal FF controls were then scaled by 0, 25, 50, 75, 100, and 125%, and the gap in reproducing individual cycle muscle activations was filled by fitting FB synergy controls as a function of joint positions, velocities, and moments as surrogates for muscle lengths, muscle velocities, and tendon forces. Finally, the six synergy-based FF+FB models controlled the subjects personalized walking model in predictive simulations performed for 3 testing walking cycles withheld from calibration. The 100% FF model (which still had minimal FB) reproduced the testing walking cycles the most closely, and only the 75%, 100%, and 125% FF models generated near-periodic walking motions using initial conditions consistent with experimental values. The 0, 25, and 50% FF models could generate near-periodic walking motions only when the initial conditions were allowed to diverge substantially from experimental values. Our findings suggest that predictive simulations of walking using real experimental data may require a minimum level of feedforward control and sufficient fitting data to predict a subjects actual dynamically consistent motion.

Conventional therapy after stroke focuses on reducing physical impairments. However, the decisions that guide peoples movements may have far-reaching consequences towards recovery. We lack the tools to characterize these decisions. Recently, researchers have created a quantitative behavioral assessment of effort-based decision-making and applied it to some clinical populations. The purpose of this paper is to examine the feasibility of evaluating effort-based decision-making during walking after stroke. We recruited five neurotypical participants in an initial study. We conducted a subjective effort valuation on the neurotypical individuals with and without a knee immobilizer to simulate the biomechanics of reduced knee flexion during post-stroke gait. Participants cleared obstacles of varying heights during overground walking, followed by rating their perceived effort and then completing an effort choice paradigm to calculate subjective effort value. In a second experiment, we recruited five individuals with stroke to perform a similar protocol without an immobilizer during harnessed treadmill walking. We found that rated perceived effort increased monotonically with obstacle height across groups, that individuals could recall obstacle heights without cues, and that subjective effort value increased with knee immobilization in the control group as expected. We conclude that adapting an effort-based decision-making assessment to a walking context in people with stroke is feasible.

ObjectiveAchilles tendon ruptures lead to long-term structural and functional deficits. Prior research that sought to identify optimal rehabilitation techniques was fundamentally limited by the inability to continuously monitor Achilles tendon loading during rehabilitation. Our objective was to develop a data-driven model that predicts per-step peak Achilles tendon loading from only a single, boot-mounted accelerometer. MethodsNineteen patients recovering from an acute Achilles tendon rupture completed in-lab walking trials while wearing an instrumented immobilizing boot. A boot-mounted inertial measurement unit provided acceleration signals used for prediction, while a force-sensing insole provided ground truth tendon-loading data through a validated ankle moment balance. We developed a stance-detection algorithm, as well as a personalized one-dimensional convolutional neural network (1D-CNN) to estimate per-step peak Achilles tendon load. Our training framework incorporated a small patient-specific personalization sample and was evaluated on held-out steps. ResultsThe stance detection algorithm identified stance phases with 99.8% precision and mean timing errors of 27.3 ms for heel strike and 61.9 ms for toe-off. The CNN estimated per-step peak Achilles tendon load with a mean absolute error of 0.14 bodyweights (R2=0.68) across rupture patients. ConclusionContinuous, objective estimation of Achilles tendon loading during early rehabilitation is feasible using a single, boot-mounted accelerometer. Model errors were small (9%) relative to the wide range of tendon loading exhibited during immobilizing boot walking. Our proposed approach enables clinicians to continuously monitor mechanical loading during a previously unobservable rehabilitation period and provides a foundation for personalized rehabilitation guidance after Achilles rupture.

BackgroundInertial measurement units (IMUs) have potential to be inexpensive, portable sensors for collecting gait parameters and joint kinematics. Current validation protocols generally do not investigate IMU accuracy in measuring altered gait; therefore, they cannot assess an IMUs ability to detect pathologies. The Stridelink IMU-based gait analysis device is intended for use in detecting and monitoring gait abnormalities, thus there is a need to evaluate the devices accuracy under abnormal gait conditions. Research questionHow well do measurements from the StrideLink IMU agree with motion capture (MoCap), particularly when gait is mechanically altered to simulate pathology? MethodsTwenty-eight healthy participants (ages 18-40) were analyzed during a one-minute tread-mill walk with Vicon MoCap and StrideLink. Tests were performed under normal and mechanically induced abnormal conditions (knee brace, walking boot). Equivalence testing and correlation analysis evaluated StrideLinks accuracy against MoCap. ResultsStrideLink showed statistical equivalence (within 5%) for average cadence, stride, swing, and stance times but not double support time. Many metrics were statistically equivalent (p < .001) despite induced abnormalities. Correlation analysis showed almost perfect agreement with MoCap for stride times, cadence, and stance. However, the abnormal gait protocol revealed nuances not observed in normal gait; specifically, the device underestimated swing time by [~]10 ms in knee brace restricted limbs. SignificanceThis study utilized mechanically induced gait abnormalities to assess the robustness of IMU measurements. Results indicate StrideLink yields valid temporal gait measurements comparable to reference systems, even under conditions of significant deviation, supporting the utility of using induced abnormalities for sensor validation.

BackgroundCommercially-available microprocessor-controlled prosthetic knees are unable to fully replicate the biomechanical function of the missing biological limb. While powered prostheses have the capacity to restore joint level kinetics, current systems rely on intrinsic control schemes that do not allow the user to volitionally modulate movement under neural commands. This limitation may compromise functional performance and hinder prosthetic embodiment, the sense that the device is part of the users body. In a case study on one test participant, we evaluate the functional and perceptual benefits of a bone-anchored, neurally-controlled knee prosthesis by comparing it to the participants microprocessor-controlled prosthesis. MethodsWe conducted a within-subject study on an individual with a transfemoral amputation, with an osseointegrated implant and surgically reconstructed agonist-antagonist muscle pairs. We tested a neurally-controlled powered knee and conventional microprocessor knee across a set of activities, including seated volitional control tasks, sit-to-stand transitions, squatting, level-ground walking, stair ascent, and uninstructed standing. Performance metrics included knee kinematics, prosthesis-generated mechanical power, and functional outcomes such as gait speed, stair ascent time, and weight-bearing symmetry derived from ground reaction forces. Functional mobility and control were complemented by self-reported embodiment, assessed through a questionnaire targeting agency, ownership, and body representation. ResultsThe neurally-controlled prosthesis enabled intuitive and responsive control. Compared to the subjects prescribed prosthesis, the prosthesis yielded improved temporal gait symmetry during walking (symmetry index: 0.93 vs. 0.59, with 1 indicating perfect stance time symmetry), increased prosthetic-side weight-bearing during sit-to-stand and squatting, and successful execution of a step-over-step stair ascent strategy--an outcome not achievable with the subjects prescribed device. Embodiment scores were consistently higher with the neurally-controlled prosthesis compared to the prescribed device across multiple domains, including agency, ownership and body representation. ConclusionsThis study is the first to directly compare a prescribed microprocessor knee with a bone-anchored, neurally-controlled powered prosthesis. By combining osseointegration, surgically reconstructed agonist-antagonist muscle pairs, and powered actuation, the system improved gait symmetry, greater prosthetic-side loading, and step-over-step stair ascent. These results demonstrate the novelty and promise of integrating surgical and mechatronic innovations to restore both functional mobility and embodied control after transfemoral amputation. Trial registrationThis study was approved by the Institutional Review Board at MIT (Protocol No. 2503001589).

Brain-computer interfaces (BCIs) translate brain signals into commands for external devices, with motor imagery (MI) BCIs decoding imagined movements to aid neurorehabilitation. Although high-channel EEG offers rich data, such systems are bulky and impractical for everyday use. This study assesses whether a low-channel, consumer-grade headset (Muse) can match a clinical-grade system (OpenBCI) in classifying lower limb MI and motor execution (ME). Six healthy volunteers performed left and right knees and ankles MI and ME tasks while EEG was recorded concurrently from both devices. Signals were band-pass filtered (8-30 Hz), segmented into overlapping one second windows, and features were extracted across time, frequency, and time-frequency domains. Feature dimensionality was reduced via mutual information-based minimum redundancy maximum relevance and principal component analysis. Five classifiers (support vector machine, linear discriminant analysis, k nearest neighbors, random forest, and AdaBoost) were applied to nine binary discrimination scenarios and evaluated with 10-fold cross-validation via 100 Monte Carlo iterations. Frequency domain features, particularly those derived from Welchs power spectral density, were most frequently selected. Mutual information analysis indicated that C3 and C4 electrodes were most informative for OpenBCI, while in Muse, the channels contributed more evenly, except in laterality classification scenarios, where TP9 played a key role. OpenBCI outperformed Muse in classifier-based accuracy with superiority ranging from 0.4% to 4.8%, while task-based differences were more variable, ranging from -0.3% to 8.7%. Despite its lower spatial resolution, the Muse system achieved competitive performance, especially in motor vs. rest tasks, and shows promise as an affordable, user-friendly alternative for home-based neurorehabilitation BCIs.

Intracortical brain-computer interfaces (BCIs) are used to decode motor intent from neural population activity; their main clinical application is to restore function for individuals with motor or communication deficits. However, when trying to reconstruct movement trajectories, such as in computer cursor control, even state-of-the-art decoders fall short of able-bodied performance during online BCI control. This calls for alternative approaches to improve the usability of motor BCIs. Here, we leveraged an error signal, i.e. a neural correlate of faulty motor control that can be detected from neural activity. By detecting this error signal in parallel to performing movement decoding, it is possible to perform error modulation, i.e. real-time error detection and correction during a closed-loop motor BCI task. We analyzed data from four individuals with upper limb impairment due to cervical spinal cord injury who each used an intracortical BCI to perform a continuous cursor control task with visual feedback. A classifier was trained to detect the error signal and was used to perform online error detection during BCI control to limit ongoing errors (defined as movement of the controller away from its target) without requiring any specific action from the participants. Our contribution is three-fold. First, we show that the error signal has a pre-error component. Cortical activity was already significantly modulated before the onset of the kinematically-defined error, theoretically allowing for earlier detection. Second, we show that error modulation significantly improves performance during online BCI control of cursor kinematics. Finally, we show that the error signal can be robustly leveraged across contexts, as error modulation improves performance in more complex motor tasks (involving for instance grasp and drag actions) or other environments without task-specific calibration. Overall, our results suggest that the error signal can be robustly disentangled from motor intent in cortical activity, and that even a simple linear classifier can enable error modulation in parallel to a continuous kinematic decoder, yielding more reliable and accurate BCI control.

Prosthetic devices balance functionality and usability to support activities of daily living (ADLs). However, many designs rely on rigid end effectors that, while anthropomorphic in form, lack biomimetic design principles. This mismatch increases cognitive and physical burden, reducing adoption rates. We developed the Human-inspired Actuator Modeling and Reconstruction (HAMR) process, a user-centered framework informed by individual morphology and functional needs, to generate customized agonist/antagonist tendon-actuated end effectors. Using HAMR, we created the Tendon Actuated Prosthetic Hand (TAPH), which integrates human-derived geometry with adaptive force distribution to promote natural object interaction. In a study with 12 participants without limb difference, TAPH was compared to a structurally similar tendon-actuated hand with generalized anthropomorphic geometry across three ADL tasks of varying complexity. TAPH significantly improved task performance and reduced physical effort, mental workload, and frustration, particularly during gross motor tasks. For fine motor tasks, performance improved under stable conditions but not during tasks requiring dynamic precision and continuous coordination. These findings highlight the functional benefits of biologically informed prosthesis design and support biomimetic principles in enhancing performance and user experience.

Motor imagery (MI) brain-computer interfaces (BCIs) are promising technologies for neurorehabilitation. In this context, deep learning (DL) models are increasingly being used to decode the mental imagination of movement. However, countless studies across multiple domains have shown that DL models are susceptible to bias, which can lead to disparate performance across subpopulations in terms of protected attributes, such as sex. The reported presence of sex-related information in electroencephalography (EEG) signals, widely used for MI-BCI, further raises warnings in this regard. For this reason, we conducted an in-depth analysis of the performance of DL in terms of the sex and other potential confounding factors. While an initial basic stratified analysis in terms of sex showed differences in favor of the female population, further analysis revealed that performance disparities were actually primarily driven by the discriminability of EEG patterns themselves, and not by the DL model. Moreover, DL models improve overall performance as well as per-group performance, particularly helping subjects with less discriminable EEG patterns. Our work highlights the benefits of DL methods for MI-BCI as well as the need for careful analysis when it comes to bias assessment in complex settings where multiple variables interact. We argue that in-depth studies of model behavior beyond standard performance metrics, should become widespread in the community in order to ensure the development and later deployment of fair BCI systems.

Wearable devices for gravity balancing have high potential for impact across domains, including neuromotor rehabilitation and occupational systems. Devices made from compliant mechanisms, optimized to achieve specific compensation moments at target joints, have proven effective, but thus far have solely been optimized towards gravity compensation and not other wearability criteria. In this work, we propose a multi-objective optimization framework, based on particle swarm optimization, to design a soft, gravity balancing shoulder orthosis, while taking into account wearability constraints such as undesired loading directions and device size. Using this custom framework, we pursued multiple stages of orthosis design and optimization, selecting multiple solutions to be translated to real-world prototypes. These solutions were realized via 3D printing with thermoplastic polyurethane and evaluated for mechanical performance on benchtop and in-vivo. In-vivo testing on 6 healthy individuals demonstrated relative reductions in muscle activity for the anterior deltoid and upper trapezius, by 53 % and 71 % respectively when operating the orthosis for static tasks within functional shoulder ranges of motion. Changes in muscle activation were also were observed across other muscles, including the posterior deltoid, as well as in dynamic tasks at different speeds.

After amputation, advanced prosthetic limbs offer a promising means of restoring motor function. However, state-of-the-art prostheses often rely on aggregate electromyogram (EMG) signals to decode motor intention, which limits their ability to replicate natural limb movements. Decomposing EMG signals into individual motor unit components has shown potential for more natural control, but distinguishing between individual units can be challenging when nearby signals overlap. This study demonstrates that muscle target reinnervation surgeries can naturally increase physical separation between motor unit signals, thereby mitigating this overlap. Reinnervation of individual motor units is evaluated in a rodent hindlimb model after direct nerve-to-muscle implantation. Histological and electrophysiological analyses reveal that structural changes following reinnervation surgery result in beneficial motor unit signal changes, particularly improving spatial separation between motor unit signals compared to those in intact muscle. This spatial separation contributed to fewer instances of complex, overlapping signals in reinnervated muscle recordings. Motor unit signals were leveraged to provide a proof-of-concept of precise control of a virtual prosthesis for the first time after direct nerve-to-muscle implantation surgery. These findings highlight the potential of reinnervated muscle targets as key biological interfaces that facilitate motor unit separation, reducing the burden on decomposition algorithms and improving prosthetic control.

Electroencephalography (EEG) serves as a fundamental tool in modern neurology, cognitive neuroscience, and brain-computer interfaces, but its practical application is often compromised by artifacts. Physiological artifacts are particularly intractable due to overlapping spectral features with neural signals, hindering reliable EEG interpretation. In this work, we propose Grid-based 3D Convolution-Transformer (G3DCT), an interpretable deep learning framework for EEG artifact identification. The framework embeds multi-channel EEG signals into fixed grids to leverage electrode spatial topology, employs parallel multi-branch temporal convolutions and Transformers to handle complex artifacts, and incorporates an attention module to visualize scalp activation patterns, which enhances physiological interpretability. Our evaluation on three datasets demonstrates that G3DCT outperforms existing state-of-the-art models. For challenging combined artifacts, it secures a gain of 2.8% in F1-score over the second-best model. These results demonstrate that G3DCT provides an efficient and robust solution for EEG artifact identification, which has the potential to enhance the reliability of EEG-based applications in practice.