Ziyun Ding

Ground reaction force (GRF) trajectories capture kinetic alterations in stance-phase loading, propulsion, and mediolateral control. This study presents GaitScreen, a GRF-based framework for post-acquisition prioritisation of already recorded healthy-control and pathological-gait trials. GaitScreen converts vertical, anterior–posterior, and mediolateral stance-phase GRF trajectories into a review-prioritisation score and a threshold-dependent review flag. The framework combines a multiscale temporal backbone, a convolutional neural network component, and a logistic regression component to integrate multiscale force patterns, learned convolutional representations, and raw-waveform linear information. It was evaluated using the GaitRec dataset, comprising healthy controls and individuals with lower-limb orthopaedic conditions. Under a subject-exclusive evaluation protocol, GaitScreen achieved consistent trial-level performance across ten matched runs, with an accuracy of 0.8766 ± 0.0025 , a macro-F1 score of 0.8757 ± 0.0024 , and an AUROC of 0.9425 ± 0.0015 . A high-sensitivity operating mode reduced the false-negative rate from 0.1736 to 0.1073, with an increase in false-positive review flags. The full hybrid model had a serialized size of 1.122 MB and a mean CPU inference latency of 0.626 ± 0.010 ms per sample after GRF preprocessing. The compact model size and low inference latency support computationally scalable post-acquisition processing of standardised GRF trials.

Human gait is a widely studied motor behavior, yet direct observation of skeletal muscle dynamics during walking remains limited. We present a novel multimodal gait dataset that integrates real-time B-mode ultrasound imaging with synchronized motion and surface electromyography (EMG) data. Twenty-six healthy adults (13 male, 13 female) completed walking trials across a range of conditions: three self-selected speeds (self-fast, self-paced, self-slow) and eight auditory-cued pacing conditions (60–130 beat per minute (bpm) in 10 bpm increments). Whole-body motion and ground reaction forces (GRFs) during gait were captured using a 3D motion capture system and two force plates, with a standardized 36-marker lower-body marker set. Surface EMG was recorded bilaterally from the tibialis anterior (TA), soleus, medial gastrocnemius, and lateral gastrocnemius muscles. Ultrasound probes were secured over the bilateral TA muscles to capture continuous fascicle dynamics throughout the gait cycle. Based on the motion capture data, inverse kinematics and inverse dynamics analyses were performed in musculoskeletal modelling software OpenSim to obtain joint kinematics and joint moments for each trial. This dataset enables the investigation of lower-limb muscle mechanics during gait, allowing exploration of the pathway from neural activation through muscle mechanics to musculoskeletal behavior.

Accurate lower limb joint torque estimation is important for gait analysis, rehabilitation planning, and the control of assistive devices. While inertial measurement units (IMUs) provide a portable alternative to laboratory-based motion capture systems, the influence of key design choices on deep learning-based joint torque estimation remains insufficiently understood. This study systematically evaluated five factors affecting lower limb joint torque estimation across five degrees of freedom: hip flexion, hip adduction, hip rotation, knee flexion, and ankle dorsiflexion. Using a public dataset of 20 healthy adults during treadmill walking, we examined model architecture, sampling frequency, sliding window size, IMU sensor configuration, and data normalisation method. Each configuration was evaluated over 10 independent training runs. Recurrent neural networks outperformed the other architectures, with LSTM achieving the strongest overall performance. Among the tested settings, 100 Hz sampling, a 64-sample sliding window, a four-sensor configuration, and Z-score normalisation gave the best overall results. The resulting benchmark configuration achieved an average Pearson correlation coefficient (R) of 0.97 across five lower limb joint torque estimation tasks. These findings provide an evidence-based reference for IMU-based joint torque estimation and support the development of wearable biomechanics for clinical gait analysis, rehabilitation, and assistive technology.

Objective: Amputation is a life-changing condition that leads to impaired gait, chronic pain, and progressive musculoskeletal deterioration. Functional electrical stimulation (FES) has the potential to improve mobility and manage pain; yet, clinically, prescribing this therapy largely depends on clinician experience, which limits its efficacy. This study aimed to design and evaluate the first FES gait rehabilitation therapy for individuals with limb amputation using musculoskeletal modelling techniques. Methods: We integrated musculoskeletal modelling with optimal control simulations to identify stimulation strategies capable of modulating joint loading during gait. An in-silico trial was conducted to determine target muscles for stimulation, followed by an in-vivo gait assessment to evaluate FES-assisted walking performance. Results: The in-silico study showed that stimulating the vasti muscles in the amputated limb provided mechanical advantages by modulating internal loading. Experimental gait analysis further confirmed these findings, demonstrating that the model-driven design produced favourable biomechanical effects, including reduced peak knee contact forces in the intact limb and increased vasti muscle forces in the amputated limb during FES-assisted walking. Conclusion: Stimulating the vasti muscles in the amputated limb can enhance gait dynamics in individuals with transtibial amputation. Significance: This study, which combines in-silico therapy design with in-vivo validation, paves the way for evidence-based digital tools to support effective FES-based gait rehabilitation and has potential to be extended to broader clinical populations with gait impairments arising from neurological or musculoskeletal conditions.

Wearable sensor-based gait analysis encompasses diverse tasks including human activity recognition, neurodegenerative disease diagnosis, and clinical event detection. Existing deep learning approaches are typically designed for specific tasks, lacking a unified framework that generalizes across different scenarios. We present the first systematic investigation of state space models for wearable sensor-based gait analysis, proposing FB-Mamba, a Forward-Backward Mamba architecture that achieves bidirectional temporal modeling with linear complexity through an adaptive fusion mechanism. We establish FB-Mamba as a general-purpose framework spanning three major task categories, including human activity recognition, disease detection, and event-level detection. Comprehensive experiments on five datasets with heterogeneous sensor modalities and sampling rates demonstrate that FB-Mamba achieves consistently competitive performance across all task categories while maintaining computational efficiency and favorable long-range dependency modeling capability.

Sit-to-stand movement is the prerequisite for other lower limb movements and determines the independent living ability of lower limb amputees. Although optimal control theory enables human motion prediction without experimental data, existing studies on sit-to-stand prediction for lower limb amputees have neither focused on transtibial amputees nor considered buttocks-seat contact. To address these limitations, this study aims to develop an optimal control-based digital computational framework for predicting the entire cycle sit-to-stand movement of transtibial amputees while analysing the biomechanical effect of prosthesis stiffness changes. The motion prediction was formulated as an optimal control problem, where the system dynamic constraints were incorporated into a biomechanical model of the unilateral transtibial amputee and the smooth Hunt-Crossley contact force model. To overcome the challenge of difficult to satisfy static equilibrium constraints at the initial and final time points of motion, a novel "free-fall method" inspired by natural motion was proposed to determine the vertical translation and pitch angle of the human model. Combining the above innovative methods with the numerically efficient direct collocation method, a digital computational framework for predicting the entire cycle sit-to-stand motion of unilateral transtibial amputees was established. After being validated by experimental data, the proposed framework was then used to analyse the effect of changes in prosthetic stiffness on the biomechanical behaviour of amputees. The predicted lower limb joint angles, ground reaction forces, and seat reaction force successfully capture key experimental features. Furthermore, the effects of changes in prosthetic stiffness on dynamic behaviour were revealed, fully demonstrating the unique advantage of predictive simulation that does not require experimental data. Our approach has great potential in multiple fields such as the design of prostheses and rehabilitation strategies, as well as motion planning for humanoid robots.

Introduction: Accurate joint kinematics estimation is essential for understanding human movement and supporting biomechanical applications. Although optical motion capture systems are accurate, their high cost, complex setup, and limited portability restrict use outside laboratory environments. This study proposes a lightweight, physics-informed neural network for real-time joint kinematics estimation using inertial measurement units (IMUs). Methods: We developed OrientationNN, which integrates orientation-based physical constraints into a compact multi-layer perceptron architecture to ensure biomechanically consistent joint kinematics estimation. The model was evaluated on a publicly available dataset and compared with a physics-based inverse kinematics framework (OpenSense) and conventional learning-based models including MLP, LSTM, CNN, and Transformer. Results: OrientationNN achieved an average joint angle estimation error below 5° during ambulatory motion and consistently outperformed OpenSense across all kinematic variables. The model required only 4.9 × 10³ FLOPs per frame and 10.8 KB of parameters, demonstrating high computational efficiency suitable for real-time applications. Conclusion: OrientationNN enables accurate and computationally efficient joint kinematics estimation from IMU data. The results highlight its potential as a cost-effective and scalable solution for wearable biomechanical and motion analysis applications.