Deep learning biomechanical models perform optimally when trained with large datasets, however these can be challenging to collect in gait labs, while limited augmentation techniques are available. This study presents a data augmentation approach based on generative adversarial networks which generate synthetic motion capture (mocap) datasets of marker trajectories and ground reaction forces (GRFs). The proposed architecture, called adversarial autoencoder, consists of an encoder compressing mocap data to a latent vector, a decoder reconstructing the mocap data from the latent vector and a discriminator distinguishing random vectors from encoded latent vectors. Direct kinematics (DK) and inverse kinematics (IK) joint angles, GRFs, and inverse dynamics (ID) joint moments calculated for real and synthetic trials were compared using statistical parametric mapping to assure realistic data generation and select optimal architectural hyperparameters based on percentage average differences across the gait cycle length. We observed negligible differences for DK computed joint angles and GRFs, but not for inverse methods (IK: 29.2%, ID: 35.5%). When the same architecture was trained also including the joint angles calculated by IK, we found no significant differences in the kinematics and GRFs, and improvements in joint moments estimation (ID: 25.7%). Finally, we showed that our data augmentation approach improved the accuracy of joint kinematics (up to 23%, 0.8°) and vertical GRFs (11%) predicted by standard neural networks using a single simulated pelvic inertial measurement unit. These findings suggest that predictive deep learning models can benefit from the synthetic datasets produced with the proposed technique.

A single inertial measurement unit (IMU) can be used with neural networks (NNs) to predict joint kinematics and kinetics. Recent studies, however, often employed virtual IMU data computed from marker-based systems to train NN models, without providing realistic performance assessments on real IMU data. To address this, we present a methodology for developing and blindly evaluating NNs predicting lower-limb joint angles and moments from a sacrum-worn IMU. NN architectures were trained on IMU data simulated from a public marker-based dataset (49 healthy adults), augmented using conditional generative adversarial networks to enhance variability. The developed NNs were blindly tested against a different dataset (seven healthy adults) of real IMU and synchronous marker-based data collected ad hoc after the NN development. The two datasets were collected in different labs using different protocols. These NNs were subsequently fine-tuned (retrained) with this dataset and re-evaluated on another real IMU data (three healthy adults) collected after fine-tuning. The NNs achieved strong predictive performance on virtual IMU data (average root mean squared error (RMSE) of 2.6±1.3° and 0.10±0.05 Nm/kg for joint angles and moments, respectively). However, performance degraded when applied to real IMU data: average RMSE of 4.5±2.0° for joint angles and 0.21±0.14 Nm/kg for moments. Fine-tuning with real IMU data improved model accuracy, recovering RMSEs to 2.6±0.8° and 0.19±0.11 Nm/kg for joint angles and moments, respectively. Overall, our performance metrics were within the reported ranges for systems employing multiple IMU sensors. This work highlights the importance of blinded assessment and fine-tuning for practical biomechanical applications.

Human activity recognition (HAR) lies at the core of digital healthcare applications that monitor different types of physical activity. Traditional HAR methods often struggle to adapt to variable-length, real-world activity data and to generalise across cohorts (e.g., from young to old cohorts). Thus, the aim of this study was to investigate HAR using wearable sensor data, with a particular focus on cross-cohort evaluation. Each dataset included two accelerometers (right thigh and lower back) sampling at 50 Hz, capturing a range of daily-life activities that were annotated using video recordings from chest-mounted cameras synchronised with the accelerometers. Neural networks were trained on young cohorts’ data and tested on old cohorts’ data. The effects of network architecture, sampling frequency and sensor location on classification performance were investigated. Network performance was evaluated using accuracy, recall, precision, F1-score and confusion matrices. The gated recurrent unit architecture achieved the best performance when trained solely on young cohorts’ data, with weighted F1-score of 0.95 ± 0.05 and 0.93 ± 0.05 for young and old cohorts, respectively, resulting in a highly generalizable method. Classification performance across multiple sampling frequencies was comparable. The thigh-mounted sensor consistently achieved higher performance than the lower back sensor across activities except lying. Furthermore, combining datasets significantly improved performance on the old cohort (weighted F1-score: 0.97 ± 0.02) due to increased variability in the training data. This study highlights the importance of network architecture and dataset composition in HAR and demonstrates the potential of neural networks for robust, real-world activity recognition across age-defined cohorts, specifically between young and old cohorts.

Purpose: Older adults and individuals with neuromuscular impairments face a high risk of falls, which can be mitigated by identifying effective stumble recovery strategies for rehabilitation. Studying stumble recovery through empirical methods is challenging due to injury risks and constraints on natural movement, whereas predictive neuromechanical simulations offer a viable alternative. This study aimed to use a musculoskeletal model within a predictive simulation framework to analyze human stumble recovery following anteriorly directed perturbations.Methods: Using a simplified musculoskeletal model and a reflex-based neural controller, two different scenarios for perturbations occurring in the early (20%) and late (60%) swing phases were simulated. The kinematics of the swing leg, including hip, knee, and ankle joint angles were analyzed for similarity to real human stumble recovery. Additionally, recovery strategies were identified by tracking the swing leg’s toe trajectory following perturbation.Results: Early swing perturbations elicited an elevating strategy, increasing hip and knee flexion to clear the obstacle, while late swing perturbations triggered a lowering strategy, rapidly placing the foot to restore stability. Minor deviations from experimental data were observed, particularly in ankle dorsiflexion and swing phase duration.Conclusion: This study highlights the effectiveness of predictive neuromechanical simulations in analyzing stumble recovery. The framework successfully replicated key recovery mechanisms, demonstrating its potential for rehabilitation, assistive device design, and fall prevention strategies aimed at enhancing mobility and reducing injury risk in vulnerable populations.

A major concern, often resulting in serious injuries and loss of independence, is caused by falls due to stumbling. Previous works, investigating stumble recovery responses isolated to joint or muscle responses, lacks a comprehensive view of coordinated whole-body recovery. A scoping review, on Web of Science, Scopus and PubMed, was conducted using keywords related to recovery from anteriorly-directed stumbles induced by physical obstruction of the swing limb during healthy human walking. Extracted data were experimental cohort, swing phase during stumble, perturbation apparatus, recovery strategies, biomechanical responses including joint kinematics, kinetics and muscle activity. A narrative synthesis was provided to summarize findings across studies. 23 of 337 identified studies met inclusion criteria. Limited focus on older adults or comparing them to young adults were found. Rigid and stationary obstacles were commonly used, while compliant/movable obstacles were rarely examined, limiting ecological validity. Recovery strategies varied by perturbation timing: foot elevation and lowering in early/mid-swing and late swing stumbles, respectively. Rare responses like delayed lowering and reaching were underexplored. Sagittal plane lower-limb kinematics were frequently reported, with often missing trunk motion, joint moments and EMG from proximal muscles. This review highlights the importance of rapid and coordinated whole-body responses in stumble recovery, particularly involving the support limb and trunk. Findings suggest rehabilitation should target hip extensor strength, trunk control and support limb engagement to improve recovery. Future research should investigate underexplored strategies, trunk and upper-limb involvement and age-related adaptations to enhance intervention design.