This paper proposes a scheduling optimisation framework GCN-EDQNet that integrates multi-scale feature fusion, bidirectional edge detection, and graph convolutional network (GCN). Data centre resources are modelled as a graph over distributed nodes; a multi-scale module builds hierarchical representations to capture spatial heterogeneity in resource distributions. A bidirectional edge-detection subnetwork then identifies scheduling-sensitive regions and produces an edge heatmap - assigning higher weights to edges connecting nodes with high load variation - that guides the GCN to prioritise structurally salient, mutation-prone areas. This explicit weighting mechanism enables the GCN to focus on bottleneck-prone regions and improve structure-aware feature learning. Finally, a reinforcement learning strategy enables adaptive task allocation and migration. Experiments on two public datasets show that GCN-EDQNet outperforms conventional approaches in task completion time, load variance, scheduling success rate, and energy efficiency. These results highlight a structure-aware, intelligent paradigm for data centre resource scheduling with clear theoretical and practical value.
The current wearable motion tracking devices have significant differences in heart rate monitoring, inaccurate calorie tracking, and measurement errors in exercise speed and distance. Based on this, this paper optimises the design of wearable motion tracking devices. Firstly, this paper establishes a real-time data mining model for sports training wearable devices using fuzzy algorithms, and determines the heterogeneity of sports training data. Then, this paper constructs a feature extraction model and processes the data using thresholding and Savitzky Golay filtering. Subsequently, this paper elaborates on the calibration method of sensors in wearable motion tracking devices, and finally tests the application of the device in training monitoring and evaluation. The research results indicate that the heart rate of student 11 measured by the device in this paper is 84 beats per minute under normal conditions and 123 beats per minute under high-intensity exercise.
This paper proposed an attention-based multi-scale deformation prediction network (AMSD-Net) for nonlinear mechanical response modelling. Using multi-dimensional physical parameters of pipeline steels as inputs, AMSD-Net integrates a hierarchical feature extraction backbone composed of Inception modules, squeeze-and-excitation (SE) channel attention, and convolutional block attention module (CBAM) spatial attention to capture deformation characteristics at different spatial scales. Parallel multi-scale convolutional pathways and a dual-attention mechanism are employed to recalibrate channel-wise and spatial features in a data-driven manner. Experimental evaluations on simulation datasets generated from X70 and X90 pipeline steels show that AMSD-Net achieves lower root mean square error and mean absolute error in stress, strain, and deformation prediction compared with representative baseline models, while maintaining stable fitting behaviour across the elastic-plastic transition region. AMSD-Net outperforms conventional baselines in predicting nonlinear deformation and failure strength, enabling more efficient and accurate data-driven pipeline integrity assessment.
This study adopts modality-specific feature extraction for text, visual, and audio inputs. Task predictions and modality representations are embedded into an adaptive graph, which is further augmented by introducing an attenuated higher-order common-neighbour similarity matrix within a heterogeneous graph neural network. This formulation is used to guide node aggregation and to support interpretability through explicit graph-based relational modelling. Based on these components, an attention-aware graph embedding model is constructed for downstream analysis. Across the Alibaba and IMDB datasets, the proposed method achieves average gains of 6.13% (Macro-F1) and 6.57% (Micro-F1) over graph embedding baselines. On IMDB, it further improves accuracy by 4.1%, F1-score by 5.9%, and reduces mean absolute error by 6.2%. These results suggest that the proposed graph-based fusion strategy can provide measurable gains on the considered benchmarks while enabling adaptive estimation of inter-modal interaction weights.
The rapid expansion of online education has made the analysis of users' implicit behaviours - viewed through the lens of nonlinear and complex data - a crucial avenue for enhancing educational effectiveness. To address this, we introduce a random forest-fuzzy comprehensive evaluation (RF-FCE) method embedded within a clustering framework. Leveraging multiple clustering techniques, we first identify distinct category-specific influence patterns across different courses. Subsequently, we integrate fuzzy comprehensive evaluation with machine learning to analyse implicit behavioural data, examining both the intrinsic factors that affect course outcomes and the complex interactions between these factors and course quality. Our findings reveal significant variations in user engagement and learning outcomes across courses of differing quality, with these variations exerting a substantial influence on learning behaviours. In summary, this study offers a structured and robust analytical approach for examining implicit user behaviours in online education, demonstrating both methodological innovation and practical utility for improving course design and delivery.
In the current diagnosis of cancer, the analysis of pathological section images and molecular markers (such as HER2, hormone receptor status, etc.) is usually performed separately, which can easily lead to difficulties in early identification, deviations in subtype classification, and limitations in personalized treatment decisions. This research solves this problem by establishing a breast cancer diagnosis model based on visual converter (ViT) and full connected neural network (FCNN). The experimental results show that the diagnostic model established in this study performs the best in terms of accuracy (0.963), recall rate (0.947), precision (0.952), and F1 score (0.950). In addition, the model shows high accuracy in classifying eight breast cancer subtypes in the cancer histopathological image dataset. The diagnostic model established in this study is helpful in promoting the development of precision medicine for cancer, improving the efficiency of clinical treatment, and has important practical value in reducing cancer mortality.
Integrated genomic and transcriptomic analyses were employed to identify novel molecular targets and elucidate underlying mechanisms in the progression from unruptured to ruptured intracranial aneurysms (IAs). The study integrated differentially expressed genes identified through GWAS-based SNP screening and transcriptomic analysis of three independent datasets (GSE13353, GSE26969, GSE75436). Protein-protein interaction network construction and functional enrichment analysis of overlapping genes revealed two key interactions: ITGAX-JAM3 and KLHL28-TOGARAM1, with ITGAX and MAGI2 emerging as consensus genes across all datasets. Machine learning-based prioritisation via LASSO regression with L1 penalty selected optimal characteristic genes, validated through ROC curve analysis (AUC > 0.85). These findings demonstrate that ITGAX, JAM3, KLHL28, TOGARAM1, and MAGI2 represent promising molecular targets worthy of further investigation in the context of hypertension-driven mechano-immune crosstalk during IA rupture progression, providing new insights for both mechanistic studies and clinical management strategies.
Skin cancer is one of the most common types of cancer worldwide, with increasing incidence rates over the past few decades. This paper proposes an optimisation-enabled deep learning for skin cancer segmentation and detection. Initially, the input image is pre-processed by a bilateral filter. After that, skin lesion segmentation is performed using LadderNet, which is tuned by Stock exchange trading white shark optimisation (SEWSO). Here, SEWSO is the combination of Stock exchange trading optimisation (SETO) and white shark optimisation (WSO). Moreover, segmented image is allowed through data augmentation which is done by rotation, shifting and random brightness techniques. Thereafter, the feature extraction is achieved to obtain the desired features. At last, the feature vector is subjected to skin cancer detection, which is accomplished by employing SqueezeNet tuned by SEWSO. This approach delivered high accuracy, sensitivity, and specificity of 93.90%, 95.00%, and 94.70%, respectively.
This study aims to investigate the role and prognostic value of SP100 in the tumour immune microenvironment of HNSCC. A comprehensive bioinformatics analysis was conducted. We first investigated the expression and overall survival (OS) of SP100 in pan-cancers, then, the relationship between tumour immune microenvironment of SP100 in HNSCC was analysed, the SP100-related genes were identified, and finally, a new gene signature was established. SP100 revealed differential expression, correlated with OS in various cancer types, and showed positive associations with immune cell infiltration and immune checkpoints. KEGG analysis showed a focus on 'antigen processing and presentation' and 'natural killer cell-mediated cytotoxicity'. Key SP100-related genes are SP110, MT2A, and CAV1. High SP100 expression negatively correlated with prognosis and positively associated with tumour immune infiltration in HNSCC. Thus, SP100 may be served as a new prognostic biomarker as well as a new target for immune therapy in HNSCC.
In the world, liver and lung cancer are the two types of cancers that occur in the human body. Liver and lung tumour segmentation is a basic process in treating and diagnosing diseases. The automated detection of these two cancers undergoes stages like dataset collection, pre-processing, and optimisation-based segmentation. Datasets like 3DIRCADb for the liver and LIDC-IDRI for the lung are initially obtained. Then, the hybrid deep learning model with optimisation is carried out for the segmentation process. The deep learning model attention encoder decoder-based residual-UNet is used to segment the liver and extract the region of interest. Similarly, the same process is carried out for lung tumour segmentation. The metaheuristic optimisation fire hawk algorithm is introduced. The segmentation performance of the proposed liver and lung segmentation model is carried out using different measures. On the liver and lung datasets, the proposed approach achieves dice values of 0.901 and 0.916, respectively.
This study employs single-cell RNA sequencing (scRNA-seq) to analyse oesophageal squamous cell carcinoma (ESCC), identifying 10 potential biomarkers (ALDH2, ANGPT2, APPL1, ARPC2, CAD, CALM1, CLDN7, CLTB, F2RL3, LPAR1) associated with radiation exposure. Methodology involves scRNA-seq for data partitioning, pre-processing, clustering, and differential expression analysis. Dysregulated genes are identified through comprehensive gene ontology (GO) annotations, and ESCC-related pathways are explored via the Kyoto encyclopedia of genes and genomes (KEGG) database. Analysis of 38 genes reveals distinct patterns under radiation exposure, enriching understanding of ESCC-related processes, components, and functions. This research provides a holistic view of ESCC's molecular landscape, emphasising the clinical significance of identified biomarkers and contributing significantly to the understanding of this complex malignancy.
A pandemic caused by a virus known as COVID-19 has swept across the globe. One potential weapon in the fight against COVID-19 could be early detection through the use of chest X-ray images. In this paper, I have used modified VGG-16 deep learning model for binary classification of COVID-19 chest X-ray images. There are 16 weight layers in the standard VGG-16 model. In the suggested modified VGG model, the total number of weight layers has been reduced from 16 to 9 (eight convolutional layers and one fully connected layer). According to the results, the modified VGG-16 model performs better than the other three models (CNN, KNN and VGG-16) in terms of quantitative measures of accuracy, sensitivity and specificity. The dataset used for the proposed work consists of 24,000 chest X-ray images of lung collected from online depository comprising of 12,000 for each class (healthy and pneumonia).
