Crowd flows prediction is an important problem for traffic management and public safety. Graph Convolutional Network (GCN), known for its ability to effectively capture and utilize topological information, has demonstrated significant advancements in addressing this problem. However, GCN-based models were often based on predefined crowd-flow graphs via historical movement behaviors of human beings and traffic vehicles, which ignored the abnormal changes in crowd flows. In this study, we propose a multi-scale fusion GCN-based framework with Tucker decomposition named mTDNet to enhance dynamic GCN for crowd flows prediction. Following the paradigm of extant methods, we also employ the predefined crowd-flow graphs as a part of mTDNet to effectively capture the historical movement behaviors of crowd flows. To capture the abnormal changes, we propose a Tucker decomposition-based network with the product of the adjacency matrix of historical movement pattern graphs and an Adaptive Learning Tensor ( ALT ) by reconstructing the crowd flows. Particularly, we utilize the Tucker decomposition scheme to decompose ALT , which enhances the dynamic learning of graph structures, allowing for effective capturing of the dynamic changes in crowd flow, including abnormal changes. Furthermore, a multi-scale 3DGCN is utilized to mine and fuse the multi-scale spatio-temporal information from crowd flows, to further boost the mTDNet prediction performance. Experiments conducted on two real-world datasets showed that the proposed mTDNet surpasses other crowd flow prediction methods.

Sleep stage classification is a critical concern in sleep quality assessment and disease diagnosis. Graph network based studies for sleep stages classification have achieved promising performance. However, these studies still ignored the importance of learning morphological feature information with the spatial-temporal relationship among multi-modal physiological signals. To address this issue, we propose a Spatial-temporal Adaptive Graph Network named StAGN for sleep stage classification. The main advantage of StAGN is to adaptively learn the time-dependent and channel-wise interdependent waveform morphological features in multimodal physiological signals. Such features will be extracted by a modified 1-dimensional ResNet with a projection shortcut connection and adjusted by a joint spatial-temporal attention, thereby best serving the followed brain topological connection graph network for sleep stage classification. Meanwhile, we leverage the contrastive learning scheme with label information to further improve classification accuracy without changing the signal morphology. Experiment results on two publicly available sleep datasets of ISRUC-S1 and ISRUC-S3 show that the proposed StAGN can achieve a competitive performance for sleep stage classification, which is superior to the state-of-the-art counterparts.

Fall is a major threat to stroke survivors with the problems of gait and balance disorders in the rehabilitation phase following severe consequences on quality of life and a heavy burden to their families. Many solutions have been proposed to assess fall risk for elders based on inertial sensor-based signals, however, there still exists a great challenge of transferring them from elderly populations to the stroke-survivors populations as gait disorder patterns are significant difference between elders and stroke survivors. In this study, we conduct a pilot study to collect inertial sensor-based signals from stroke survivors when they performed the timed up and go test, and build an automatic fall risk assessment model with the architecture of Siamese network, with a merit of mitigating the problem of small sample size. Specifically, the proposed automatic fall risk assessment model consists of two parallel convolutional neural networks, each of which is composed of three convolutional layers, two max-pooling layers, and three fully connected layers. To utilize the space relation among accelerator-based and gyroscope-based signals, two-dimensional discrete wavelet transform extracts image-like features, wavelet coefficients, from inertial sensor-based signals as the input. Experimental results show that the proposed fall risk assessment model has achieved a promising results, which outperform cutting-edge methods with a big margin. The proposed fall risk assessment model with low computational complexity and limited memory consuming can be deployed on an embedded system to provide fall risk assessment service for stroke survivors in point-of-care environments or community settings.