Objective. Deep learning (DL) models have been proven to be effective in decoding motor imagery (MI) signals in Electroencephalogram (EEG) data. However, DL models’ success relies heavily on large amounts of training data, whereas EEG data collection is laborious and time-consuming. Recently, cross-dataset transfer learning has emerged as a promising approach to meet the data requirements of DL models. Nevertheless, transferring knowledge across datasets involving different MI tasks remains a significant challenge in cross-dataset transfer learning, limiting the full utilization of valuable data resources. Approach: This study proposes a pre-training-based cross-dataset transfer learning method inspired by Hard Parameter Sharing in multi-task learning. Different datasets with distinct MI paradigms are considered as different tasks, classified with shared feature extraction layers and individual task-specific layers to allow cross-dataset classification with one unified model. Then, Pre-training and fine-tuning are employed to transfer knowledge across datasets. We also designed four fine-tuning schemes and conducted extensive experiments on them. Main results: The results showed that compared to models without pre-training, models with pre-training achieved a maximum increase in accuracy of 7.76%. Moreover, when limited training data were available, the pre-training method significantly improved DL model’s accuracy by 27.34% at most. The experiments also revealed that pre-trained models exhibit faster convergence and remarkable robustness. The training time per subject could be reduced by up to 102.83 s, and the variance of classification accuracy decreased by 75.22% at best. Significance: This study represents the first comprehensive investigation of the cross-dataset transfer learning method between two datasets with different MI tasks. The proposed pre-training method requires only minimal fine-tuning data when applying DL models to new MI paradigms, making MI-Brain–computer interface more practical and user-friendly.

Locomotion recognition and prediction is essential for real-time human–machine interactive control. The integration of electromyography (EMG) with mechanical sensors could improve the performance of locomotion recognition. However, the potential of EMG in motion prediction is rarely discussed. This paper firstly investigated the effect of surface EMG on the prediction of locomotion while integrated with inertial data. We collected EMG signals of lower limb muscle groups and linear acceleration data of lower limb segments from ten healthy participants in seven locomotion activities. Classification models were built based on four machine learning methods—support vector machine (SVM), k-nearest neighbor (KNN), artificial neural network (ANN), and linear discriminant analysis (LDA)—where a major vote strategy and a content constraint rule were utilized for improving the online performance of the classification decision. We compared four classifiers and further investigated the effect of data fusion on the online locomotion classification. The results showed that the SVM model with a sliding window size of 80 ms achieved the best recognition performance. The fusion of EMG signals does not only improve the recognition accuracy of steady-state locomotion activity from 90% (using acceleration data only) to 98% (using data fusion) but also enables the prediction of the next steady locomotion (∼370 ms). The study demonstrates that the employment of EMG in locomotion recognition could enhance online prediction performance.

Patients with early-stage Parkinson’s disease (PD) exhibit various but subtle motor symptoms, especially postural instability and gait disorders (PIGD). Patients show deteriorated gait performance at turns as the complex gait task requires more limb coordination and postural stability control, which may help to discriminate signs of early PIGD. In this study, we firstly proposed an IMU-based gait assessment model for quantifying comprehensive gait variables in both straight walking and turning tasks from five domains: respectively gait spatiotemporal parameters, joint kinematic parameters, variability, asymmetry, and stability. Twenty-one patients with idiopathic Parkinson’s disease at the early stage and nineteen age-matched healthy elderly adults were enrolled in the study. Each participant wore a full-body motion analysis system with 11 inertial sensors and walked along a path consisting of straight walking and 180-degree turns at a self-comfortable speed. A total of one hundred and thirty-nine gait parameters were derived for each gait task. We explored the factor effect of group and gait tasks on gait parameters using a two-way mixed analysis of variance. The discriminating ability of gait parameters between PD and the control group was evaluated using receiver operating characteristic analysis. Sensitive gait features were optimally screened (AUC>0.7) and categorized into 22 groups to classify PD and healthy controls based on a machine learning method. Results demonstrated that PD patients exhibited more gait abnormalities at turns, especially on the RoM and stability of the neck, shoulder, pelvic, and hip joints compared to the healthy control group. These gait metrics have good discriminating abilities to identify early-stage PD (AUC>0.65). Moreover, the inclusion of gait features at turns can significantly improve the classification accuracy compared to that only used parameters during straight walking. We show that quantitative gait metrics during turning have great potential to be used for enhancing early-stage PD detection.

Objective: Current inertial-based models were mostly limited to gait assessment of straight walking, which may not be efficient for detecting subtle gait disorders at an early stage of Parkinsons disease (PD). As PD patients exhibit more severe gait impairments during turns even before the appearance of gait disorders, gait characteristics during turning can provide promise in the identification of early-stage PD. Methods: We proposed a novel spatiotemporal gait model using inertial measurement units that can assess gait performance in both straight walking and turning. Ten healthy young, ten healthy elderly subjects and ten early-stage PD patients were enrolled in the validation experiment. All participants performed a 7-meter walk test consisting of a straight walking path and turns at a self-selected speed. Spatiotemporal gait parameters from the proposed model were compared with the Vicon motion capture system. Results: A strong correlation of all spatiotemporal parameters (Pearsons R between 0.82 ~ 0.99) between the inertial-based model and the reference was observed. Most measurement differences were within the mean 1.96 standard deviation lines. The absolute bias was below 6.21 ms for all temporal gait parameters, 2.19 cm for stride length and 0.02 m/s for walking speed. We show that the proposed model does not only achieve a highly accurate and reliable spatiotemporal gait measurement but also enable the detection of significantly decreased stride length and reduced walking speed in early-stage PD patients at turns compared to the control groups. Significance: Our model offers a potential approach for early-stage PD detection.

To maximise the efficiency of gait interventions, gait phase and joint kinematics are important for closing the system loop of adaptive robotic control. However, few studies have applied an inertial sensor system including both gait phase detection and joint kinematic measurement. Many algorithms for joint measurement require careful alignment of the inertial measurement unit (IMU) to the body segment. In this paper, we propose a practical gait feedback method, which provides sufficient feedback without requiring precise alignment of the IMUs. The method incorporates a two-layer model to realise simultaneous gait stance and swing phase detection and ankle joint angle measurement. Recognition of gait phases is performed by a high-level probabilistic method using angular rate from the sensor attached to the shank while the ankle angle is calculated using a data fusion algorithm based on the complementary filter and sensor-to-segment calibration. The online performance of the algorithm was experimentally validated when 10 able-bodied participants walked on the treadmill with three different speeds. The outputs were compared to the ones measured by an optical motion analysis system. The results showed that the IMU-based algorithm achieved a good accuracy of the gait phase recognition (above 95%) with a short delay response below 20 ms and accurate angle measurements with root mean square errors below 3.5° compared to the optical reference. It demonstrates that our method can be used to provide gait feedback for the correction of drop foot.

Walking-assistive devices require adaptive control methods to ensure smooth transitions between various modes of locomotion. For this purpose, detecting human locomotion modes (e.g., level walking or stair ascent) in advance is crucial for improving the intelligence and transparency of such robotic systems. This study proposes Deep-STF, a unified end-to-end deep learning model designed for integrated feature extraction in spatial, temporal, and frequency dimensions from surface electromyography (sEMG) signals. Our model enables accurate and robust continuous prediction of nine locomotion modes and 15 transitions at varying prediction time intervals, ranging from 100 to 500 ms. Experimental results showcased Deep-STP's cutting-edge prediction performance across diverse locomotion modes and transitions, relying solely on sEMG data. When forecasting 100 ms ahead, Deep-STF achieved an improved average prediction accuracy of 96.60%, outperforming seven benchmark models. Even with an extended 500ms prediction horizon, the accuracy only marginally decreased to 93.22%. The averaged stable prediction times for detecting next upcoming transitions spanned from 31.47 to 371.58 ms across the 100-500 ms time advances. Although the prediction accuracy of the trained Deep-STF initially dropped to 71.12% when tested on four new terrains, it achieved a satisfactory accuracy of 92.51% after fine-tuning with just 5 trials and further improved to 96.27% with 15 calibration trials. These results demonstrate the remarkable prediction ability and adaptability of Deep-STF, showing great potential for integration with walking-assistive devices and leading to smoother, more intuitive user interactions.