19 October 2023 in IEEE Transactions on Instrumentation and Measurement
This paper develops a novel distributed framework based on physics-informed deep learning for robust and efficient musculoskeletal modelling in nonstationary scenarios, which could simultaneously strengthen the robustness and generalization, and reduce the time cost of model training. Without loss of generality, we utilize surface electromyogram (sEMG)-based muscle forces and joint angle prediction as an example to demonstrate the proposed distributed framework. Specifically, the whole collected sEMG data are first divided into several sub-domains, and then corresponding number of physics-informed deep learning-based local models is built using these grouped data. Finally, all the local models are integrated into a global model to obtain the final predictions. Moreover, weights inversely proportional to the training errors of local models are added to the corresponding local models to reduce and control negative effects of unknown factors. Different from existing distributed modelling methods, the proposed distributed framework embeds the prior physics knowledge, i.e., the equation of motion, into local models to regularise loss functions of deep neural networks, it thus could overcome limitations of the conventional data-driven and physics-based musculoskeletal models while preserve their advantages. The local-global distributed modelling mechanism could locally achieve better representation for sub-domains while preserve the global performance, and reduce the computational cost and memory requirements. Additionally, the embedded prior physics knowledge enables local models to reflect physical or physiological mechanisms during the training process, which could alleviate the overfitting problem and reduce the need of the number of training data, and thus the global model is more robust and better generalizes to the unseen data. Comprehensive experiments on six healthy subjects demonstrate the feasibility and effectiveness of the proposed distributed framework.
4 November 2022 in IEEE Transactions on Instrumentation and Measurement
Various spatial filters have been proposed to enhance the target identification performance of steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs). The current methods only extract the target-related information from the corresponding stimulus to learn the spatial filter parameter. However, the SSVEP data from neighboring stimuli also contain frequency information of the target stimulus, which could be utilized to further improve the target identification performance. In this paper, we propose a new method incorporating SSVEPs from the neighboring stimuli to strengthen the target-related frequency information. First, The spatial filter is obtained by maximizing the summation of covariances of SSVEP data corresponding to the target and its neighboring stimuli. Then the correlation features between spatially filtered templates and test data are calculated for target detection. For the performance evaluation, we implemented the offline experiment using the 40-class benchmark dataset from 35 subjects and the 12-target self-collected dataset from 11 subjects. Compared with the state-of-art spatial filtering methods, the proposed method showed superiority in classification accuracy and information transfer rate (ITR). The comparison results demonstrate the effectiveness of the proposed spatial filter for target identification in SSVEP-based BCIs.
12 June 2024 in IEEE Transactions on Instrumentation and Measurement
Depressive disorder has become a common problem among higher education students, but it often gets undiagnosed and untreated due to unrecognized symptoms, poor access to medical resources, and fear of stigma. To improve the situation, automatic depression detection would be essential. In this paper, we explore the feasibility of depression detection in higher education students using their behavioral data automatically collected by the University system. First, a DeepFM network, which can not only take discrete-continuous mixed features as its input but also can learn linear and nonlinear relations between the input and the output, is presented for depression detection. A modified focal loss function (MFL) is then proposed to alleviate data imbalance impact caused by the fact that the proportion of healthy students outweighs those diagnosed with depression significantly. To verify the effectiveness of the proposed method, behavioral data from 3218 students were collected, of which 179 were diagnosed with depression by university psychologists using PHQ-9 scale scores. 5-fold cross-validations are performed, and the experiment results have illustrated that DeepFM obtains the highest average accuracy compared to Multilayer Perceptron (MLP), Factorisation Neural Network (FNN), and Product-based Neural Network (PNN), demonstrating the effectiveness of the proposed framework for depression detection among university students.
4 June 2025 in Neurocomputing