ASTGSleep: Attention-Based Spatial–Temporal Graph Network for Sleep Staging

LightGCAN: A lightweight graph convolutional attention network for user preference modeling and personalized recommendation

Vehicle trajectory prediction is an important research basis for the decision making and path planning of the intelligent and connected vehicle. In the connected vehicle environment, vehicles share information and drive cooperatively, and the intelligent and connected vehicles are able to obtain more accurate and rich perception information, which provides a data basis for accurate prediction of vehicle trajectories. However, attaining accurate and effective vehicle trajectory predictions poses technical challenges due to insufficient extraction of vehicular spatial–temporal interaction features. In this paper, we propose a vehicle trajectory prediction model based on graph convolutional neural network (GCN) in a connected vehicle environment. Specifically, using the driving scene information obtained by the intelligent and connected vehicle, the spatial graph and temporal graph are constructed based on the spatial interaction coefficient (SIC) and self-attention mechanism, respectively. Then, the graph data are entered into the interaction extraction module, and the spatial interaction features and temporal interaction features are extracted separately using the graph convolutional networks, which are fused to obtain the spatial–temporal interaction information. Finally, the interaction features are learned based on the convolutional neural networks to output the future trajectory information of all vehicles in the scene by one forward operation rather than a step-by-step process. The ablation experiment results show that the method proposed in this study to model the spatiotemporal interaction among vehicles based on SIC and self-attention mechanism reduces the prediction error by 5% and 12%, respectively. The results from the model comparison experiment show that the proposed method engenders an 8% improvement in prediction accuracy over the state-of-the-art solution, providing technical and theoretical support for trajectory prediction research of intelligent and connected vehicles.

Graph convolutional networks (GCNs) can effectively learn graph data features and are widely used to predict the bearing remaining useful life (RUL). However, most existing GCN-based methods are based on single-node relationships to construct the graph structure while ignoring potential complex relationships between nodes. To enrich graph feature information and improve bearing RUL prediction accuracy, a novel bearing RUL prediction method is proposed in this article, which use spatial–temporal information fusion graph convolution network. The proposed method mines implicit temporal dependence and spatial correlation in graph samples by constructing PathGraph and RadiusGraph, which effectively improves the feature diversity of graph samples. In addition, to efficiently extract the spatial–temporal features of graph samples and alleviate the over-smoothing problem in GCNs, a feature extraction module combining graph attention network (GAT) and graph convolution network–temporal convolution network (GCN-TCN) is designed, where GAT dynamically assigns edge weights to capture spatial dependencies, while GCN-TCN enhances global temporal dependency capture. The experimental validation results show that the proposed method effectively improves bearing RUL prediction accuracy.

The neuroimaging of humans using 7T magnetic resonance imaging (MRI) has been conducted using phased array (PA) coils with different numbers of receiving channels. PA coils with a high number of channels may offer parallel imaging (PI) with a high reduction (R)-factor, which is enabled via under-sampling and coil geometry (g) factor, increasing the radiofrequency signal sensitivity provided by a small coil. The goals of this study were to assess and validate the coil performance of PA coils with different numbers of receiver (Rx)-channels in and to propose the coil selection guidelines by visualizing 7T brain images. The combined magnetic flux density (||B1||) distributions of four configurations of PA coils—4-, 8-, 12-, and 16-channel Rx-only mode under the local transmit (Tx) mode of birdcage coils—were evaluated using electromagnetic (EM) calculations. These four configurations of PA coils and a local Tx coil were designed and built for a 7T MRI experiment. For 7T brain imaging experiments, all PA coils with (w/) and without (w/o) R-factors were compared in terms of signal-to-noise ratio (SNR) and spatial noise variation (SNV). EM simulation results clearly demonstrated that PA coils with a high number of Rx channels showed more homogeneously distributed ||B1|| fields than a PA coils with a low number of Rx coils. The results of this study demonstrate that a collection of smaller surface coils can contribute to high RF signal sensitivity in terms of the anatomical coverage of the brain and may facilitate PI. With further improvement in coil technology, researchers and clinicians will be provided with PA coils with different numbers of channels, which can ensure the optimum SNR and PI benefits for 7T brain MR imaging.

Objectives: This paper proposed a new method for performance enhancement of Optical Add or Drop Multiplexer (OADM) with the Dense Wavelength Division Multiplexing (DWDM) based on the artificial intelligence. Methods/Statistical Analysis: The FF-NN is trained and tested in the MATLAB platform from the parameters obtained from simulation. The NN is trained for 70 signals with different number of channels and with different data rates as well as it is tested with 30 signals. The classification of the signals to be added or dropped is determined by the FF-NN. The DWDM network is initially modeled with the OptiSystem software tool with different number of channels with different data rates and channel spacing. Findings: The parameters BER, OSNR, Jitter and Chromatic Dispersion have been calculated. The training and testing of the neural network is carried out on the MATLAB platform and based on this, the signal will add/drop or allow the signal to pass. The result shows that the proposed method achieves the accuracy of 97.28% on classifying the signals to be dropped from the fiber or passed through it without any interruption regarding its ability to make the transmission with minimum error. The performance of the proposed method is also analyzed in terms of Transmitted & Received Signal Power and compared with the conventional OADM system with multiple filters. Applications/Improvements: The proposed method offers a viable solution to increase the performance of OADM with the DWDM. The improvement includes the use of hybrid algorithms, to further increase the performance.

The Voxel Imaging PET (VIP) Pathfinder project intends to show the advantages of using pixelated solid-state technology for nuclear medicine applications. It proposes designs for Positron Emission Tomography (PET), Positron Emission Mammography (PEM) and Compton gamma camera detectors with a large number of signal channels (of the order of 106). For Compton camera, especially with a large number of readout channels, image reconstruction presents a big challenge. In this work, results are presented for the List-Mode Ordered Subset Expectation Maximization (LM-OSEM) image reconstruction algorithm on simulated data with the VIP Compton camera design. For the simulation, all realistic contributions to the spatial resolution are taken into account, including the Doppler broadening effect. The results show that even with a straightforward implementation of LM-OSEM, good images can be obtained for the proposed Compton camera design. Results are shown for various phantoms, including extended sources and with a distance between the field of view and the first detector plane equal to 100 mm which corresponds to a realistic nuclear medicine environment.

Edge features contain important information about graphs. However, current state-of-the-art neural network models designed for graph learning, e.g. graph convolutional networks (GCN) and graph attention networks (GAT), adequately utilize edge features, especially multi-dimensional edge features. In this paper, we build a new framework for a family of new graph neural network models that can more sufficiently exploit edge features, including those of undirected or multi-dimensional edges. The proposed framework can consolidate current graph neural network models; e.g. graph convolutional networks (GCN) and graph attention networks (GAT). The proposed framework and new models have the following novelties: First, we propose to use doubly stochastic normalization of graph edge features instead of the commonly used row or symmetric normalization approches used in current graph neural networks. Second, we construct new formulas for the operations in each individual layer so that they can handle multi-dimensional edge features. Third, for the proposed new framework, edge features are adaptive across network layers. As a result, our proposed new framework and new models can exploit a rich source of graph information. We apply our new models to graph node classification on several citation networks, whole graph classification, and regression on several molecular datasets. Compared with the current state-of-the-art methods, i.e. GCNs and GAT, our models obtain better performance, which testify to the importance of exploiting edge features in graph neural networks.

Graph theory provides a robust framework for modelling complex relationships in medical data, enhancing classification accuracy through relational learning. Unlike traditional machine learning (ML) models that treat data points independently, graph-based approaches support structural dependencies to improve feature representation. This study explores the application of Graph Attention Networks (GAT), GraphSAGE, Graph Convolutional Network(GCN) and Graph Isomorphism Network (GIN) for breast cancer classification, employing k-nearest neighbor (KNN) graphs to construct a structured dataset where nodes represent patients and edges capture feature similarities. The effectiveness of the graph-based approaches is evaluated against traditional ML classifiers, including Decision Trees, Random Forest, LightGBM, and XGBoost. Experimental results indicate that GCN, GIN, GAT and GraphSAGE are beat conventional methods, with GCN, GAT and GraphSAGE achieved 100% test-accuracy and with GIN achieved 99.42%, to confirm the percentage of accuracy, authors conducted extensive experiments, including robustness testing by reducing KNN connections, introducing noise, and shuffling train-test splits. Results demonstrate that graph-based models are significantly best than traditional ML models, and these graph-based models maintain same classification accuracy while maintaining stability under robustness tests. The findings confirm that graph-based learning provides a scalable, interpretable, and highly accurate alternative for medical classification tasks, proving its effectiveness in distinguishing between benign and malignant tumors.

During the acquisition of electroencephalographic (EEG) signals, data may be missing or corrupted by noise and artifacts. To reconstruct the incomplete data, EEG signals are firstly converted into a three-order tensor (multi-dimensional data) of shape time × channel × trial. Then, the missing data can be efficiently recovered by applying a tensor completion method (TCM). However, there is not a unique way to organize channels and trials in a tensor, and different numbers of channels are available depending on the EEG setting used, which may affect the quality of the tensor completion results. The main goal of this paper is to evaluate the robustness of EEG completion methods with several designed parameters such as the ordering of channels and trials, the number of channels, and the amount of missing data. In this work, the results of completing missing data by several TCMs were compared. To emulate different scenarios of missing data, three different patterns of missing data were designed. Firstly, the amount of missing data on completion effects was analyzed, including the time lengths of missing data and the number of channels or trials affected by missing data. Secondly, the numerical stability of the completion methods was analyzed by shuffling the indices along channels or trials in the EEG data tensor. Finally, the way that the number of electrodes of EEG tensors influences completion effects was assessed by changing the number of channels. Among all the applied TCMs, the simultaneous tensor decomposition and completion (STDC) method achieves the best performance in providing stable results when the amount of missing data or the electrode number of EEG tensors is changed. In other words, STDC proves to be an excellent choice of TCM, since permutations of trials or channels have almost no influence on the complete results. The STDC method can efficiently complete the missing EEG signals. The designed simulations can be regarded as a procedure to validate whether or not a completion method is useful enough to complete EEG signals.

Visual inspection of Polysomnography (PSG) recordings by sleep experts, based on established guidelines, has been the gold standard in sleep stage classification. This approach is expensive, time-consuming, and mostly limited to experimental research and clinical cases of major sleep disorders. Various automatic approaches to sleep scoring have been emerging in the past years and are opening the way to a quick computational assessment of sleep architecture that may find its way to the clinics. With the hope to make sleep scoring a fully automated process in the clinics, we report here an ensemble algorithm that aims at not only predicting sleep stages but of doing so with an optimized minimal number of EEG channels. For that, we combine a genetic algorithm-based optimization with a classification framework that minimizes the number of channels used by the machine learning algorithm to quantify sleep stages. This resulted in a sleep scoring with an F1 score of 0.793 for the fully automated model and 0.82 for the model trained on 10 percent of the unseen subject, both with only 3 EEG channels. The ensemble algorithm is based on a combination of extremely randomized trees and MiniRocket classifiers. The algorithm was trained, validated, and tested on night sleep PSG data collected from 7 subjects. Our approach’s novelty lies in using the minimum information needed for automated sleep scoring, based on a systematic search that concurrently selects the optimal-minimum number of EEG channels and the best-performing features for the machine learning classifier. The optimization framework presented in this work may enable new flexible designs for sleep scoring devices suited to studies in the comfort of homes, easily and inexpensively. In this way facilitate experimental and clinical studies in large populations.

Summary Monitoring carbon dioxide (CO2) saturation plume movement and pressure buildup is critical for ensuring the environmental safety of geological carbon storage (GCS) projects. High-fidelity numerical simulations provide accurate modeling of CO2 plume dynamics, but they are often computationally intensive. Recent advancements in data-driven models have enabled the rapid prediction of CO2 plume movement. By leveraging available simulation data sets, these models offer a more efficient alternative without compromising accuracy. In this study, we adopt the U-Net enhanced graph convolutional neural network (UGCN) to predict the spatial and temporal evolution of CO2 plume saturation and pressure buildup in saline aquifers. Utilizing the U-Net architecture, which incorporates skip connections, enables UGCN to capture high-level features and fine-grained details concurrently. We have created physics-based numerical simulation models that account for both CO2 injection and post-injection periods. By employing the Latin-hypercube sampling method, we generated a diverse range of reservoir and decision parameters, resulting in a comprehensive simulation database. We trained and tested the UGCN model on two different data sets, including a 2D radial reservoir model (Scenario 1) and a realistic Society of Petroleum Engineers Comparative Solution Project b (SPE 11b, Scenario 2) model to train and validate the UGCN model. The performance of the UGCN model was compared with other standard graph neural networks (GNNs), such as graph convolution network (GCN) and graph attention network (GAT). Notably, the UGCN model demonstrated robust performances on the blind testing dataset, achieving an R2 score of 0.993 and 0.989 for saturation predictions in Scenarios 1 and 2, respectively. Similarly, for pressure buildup the new model achieved an R2 of 0.989 and 0.999 for Scenarios 1 and 2, respectively. These prediction results indicate the effectiveness of the trained models in predicting the temporal and spatial evolution of CO2 saturation and pressure buildup predictions. Moreover, the prediction central processing unit (CPU) time for the deep learning (DL) models is significantly lower (0.02 seconds per case) than the physics-based reservoir simulator (on average, 10–15 minutes per case). This underscores the capability of the proposed method to provide predictions as accurate as physics-based simulations while reducing substantial computational costs.

International audience

Change detection via remote sensing data is a popular method for monitoring land cover/land use. Graph attention network (GAT) is a method that can improve the change detection performance of land cover/land use monitoring by enhancing the feature representation of remote sensing images. However, the shortcomings of connection sparsity and insufficient sample feature mining of the GAT affect the application of these methods in change detection. This paper proposes a Siamese autoencoder graph attention networks for object-oriented land cover/land use change detection via high-resolution remote sensing, which is useful for semi-supervised problem methods with poor simplicity. First, we reduce the pressure of graph network model operations and obtain multidimensional features via the growing multilevel segmentation strategy (GMSS). The adjacency matrix is established by adding strongly connected edges via the weighted difference similarity matching method (WDSM) with image objects as nodes. Second, we use the Siamese autoencoder to pretrain the node features and migrate the weight parameters to the GAT feature extraction layer. Finally, a small number of samples are selected to train the GAT and predict all nodes. The experimental results show that the average overall accuracy of the proposed method is 95.41%, and the average F1 score is 89.35%, which are at least 5.04% and 10.84% better than those of other typical methods, such as the graph convolutional network (GCN). In particular, detecting roads and bare soil is significantly better than that of other methods.

GAC-SleepNet: A dual-structured sleep staging method based on graph structure and Euclidean structure

Circular RNAs (circRNAs) are covalently closed single-stranded RNA molecules, which have many biological functions. Previous experiments have shown that circRNAs are involved in numerous biological processes, especially regulatory functions. It has also been found that circRNAs are associated with complex diseases of human beings. Therefore, predicting the associations of circRNA with disease (called circRNA-disease associations) is useful for disease prevention, diagnosis and treatment. In this work, we propose a novel computational approach called GGCDA based on the Graph Attention Network (GAT) and Graph Convolutional Network (GCN) to predict circRNA-disease associations. Firstly, GGCDA combines circRNA sequence similarity, disease semantic similarity and corresponding Gaussian interaction profile kernel similarity, and then a random walk with restart algorithm (RWR) is used to obtain the preliminary features of circRNA and disease. Secondly, a heterogeneous graph is constructed from the known circRNA-disease association network and the calculated similarity of circRNAs and diseases. Thirdly, the multi-head Graph Attention Network (GAT) is adopted to obtain different weights of circRNA and disease features, and then GCN is employed to aggregate the features of adjacent nodes in the network and the features of the nodes themselves, so as to obtain multi-view circRNA and disease features. Finally, we combined a multi-layer fully connected neural network to predict the associations of circRNAs with diseases. In comparison with state-of-the-art methods, GGCDA can achieve AUC values of 0.9625 and 0.9485 under the results of fivefold cross-validation on two datasets, and AUC of 0.8227 on the independent test set. Case studies further demonstrate that our approach is promising for discovering potential circRNA-disease associations.