Aiming at the difficulties present in current automatic sleep staging methods, a method for automatic sleep staging of EEG and ECG dual modal signals by combining U²-Net and CBAM fusion attention is proposed.Firstly, the EEG-ECG signals in the MIT-BIH public dataset used in this paper are preprocessed. Then, the U²-Net network with multi-scale feature extraction module is used to extract waveform features in EEG and ECG in parallel. Secondly, CBAM fusion attention is used to assign weights to all features. Finally, the Softmax activation function is used to classify sleep periods into six. The results show that when sleep staging is performed based on U²-Net and CBAM fusion attention models, the overall accuracy of hexaclassification using ECG single-modal signals is 80.2%, and the F1 score is 75.3%. The overall accuracy of six classifications using EEG single-modal signals was 85.8%, and the F1 score was 81.7%;The overall accuracy of the six classifications using EEG-ECG dual-modal signals was 90.4%, and the F1 score was 85.6%. This shows that the bimodal sleep staging model proposed in this paper is feasible and effective, and provides a new idea for automatic sleep staging.

Imaging genetics suggests that there is a certain degree of correlation between neuroimaging and genes, leading researchers to pay attention to the analysis of diseases using genetic variations and imaging data. In practice, clinical doctors usually have limited data availability but still aspire to employ deep learning method for real-world problems. Considering the expanding data scale and expensive annotation costs, it becomes essential to develop an unsupervised learning method capable of utilizing multimodal data. To meet these needs, a representation learning method based on multimodal tabular data with contrastive learning (MTCL) is proposed. The model leverages resting-state functional magnetic resonance imaging (rs-fMRI) and single nucleotide polymorphisms (SNP) data without requiring any labeled information. To enhance interpretability, the model first transforms rs-fMRI and SNP data into a tabular structure through a feature extraction module. Then, a multimodal tabular data contrastive learning method is employed to fuse the dataset and obtain the fused data representation. On the dataset of major depressive disorder (MDD), our proposed method effectively improves the diagnostic performance of MDD. Additionally, the MTCL method combines model attribution techniques to explore imaging and genetic biomarkers associated with MDD, enhancing the interpretability of the model and aiding researchers in understanding the mechanisms underlying the disease.

Multi-functional active peptide is a protein-derived compound that can act on multiple targets and deliver a variety of physiological effects, and has significant therapeutic effects on a variety of diseases. The existing multi-functional active peptide prediction model fails to fully consider the correlation between amino acids in the feature representation stage, which reduces the feature representation ability of the model, and the existing method adopts the strategy of converting the multi-label classification problem into multiple binary classification prediction problems, which leads to the inability of the model to consider the dependence between multiple functions of the active peptide in the prediction stage, which reduces the prediction accuracy of the model for multi-functional active peptides. In order to solve the above problems, a multi-functional active peptide prediction model based on label dependence is proposed, TCLD, which extracts the correlation between amino acids in the active peptide sequence through the Transformer encoder, and uses the ZLPR loss function to capture the dependence between multiple functions, which is used to improve the performance of the multi-functional active peptide prediction model. The experimental results show that the prediction performance of TCLD is better than that of the existing multi-functional active peptide prediction methods, which is helpful for researchers to quickly screen out multi-functional active peptide candidates with potential therapeutic value, thereby accelerating the research and development process of new drugs.

To improve the segmentation accuracy of cataract surgical instruments, an EE-DANet was constructed in this paper. EE-DANet adopts a double-branched structure: The edge branch is responsible for extracting edge features, using augmented spatial attention (ASA) to address the issue of edge information loss in traditional image segmentation; The main branch adopts a U-shaped structure, and the multi-scale feature fusion module is used in the feature extraction stage to solve the problem of multiple types of surgical instrument images and large scale transformation; in the decoder section, a strip coordinate attention (SCA) is used to globally model the image, capturing the direction and position information of the surgical instrument, and address the problem of mirror reflection of the surgical instrument. The IoU and Dice reached 75.5% and 83.9% on the Dataset_instrument dataset, respectively. The proposed model effectively improves the segmentation performance of cataract surgery instruments, and has a positive significance for the clinical diagnosis of cataracts.

In the early stages of drug discovery, deep generative models are emerging as crucial tools for molecular design. The simplified molecular input line entry system(SMILES) serves as a standard chemical representation widely used for model training and generation. However, due to the non-uniqueness and non-directionality of linear representations of molecular ring systems, existing unidirectional encoders face limitations in capturing the global semantic structure of samples and generating valid molecular rings. Therefore, a method called Chemical RWKV BERT (ChemRB) is proposed, which aims to deeply extract bidirectional information from a large amount of unlabeled data. To achieve this, two pre-training tasks are innovatively designed: ring-level feature prediction and global-span closure prediction. These pre-training tasks not only provide the model with rich contextual information but also further enhance its in-depth understanding of the structural properties of complex molecules. Experimental results show that the ChemRB model not only achieves significant performance improvements but also reaches optimal baseline performance on new molecular/sample evaluation metrics. This excellent performance fully validates the effectiveness of ChemRB in accurately capturing the inherent structural information of molecules, providing a solid empirical foundation for its application in related fields. Finally, through testing and application on EGFR inhibitors, the practical utility and broad application prospects of the ChemRB model are further validated.

Most Transformer-based object tracking models have limited extraction of target's local spatial feature information and insufficient utilization of temporal features, significantly affecting the performance of object tracking models in handling complex scenarios such as target occlusion, deformation, or scale changes. Therefore, a visual object tracking method with spatial-temporal feature enhancement and perception (STFEP) are proposed in this paper. On one hand, this method uses Transformer for the extraction and fusion of search region and temporal context features to obtain global feature information. By designing a local convolutional neural network, it extracts the target's local feature information and associates it with the target's global feature information, further enhancing the target's feature representation. On the other hand, a spatial-temporal feature perception mechanism is proposed to analyze the reliability and necessity of feature information at different moments, constructing dynamic templates to perceive richer spatial-temporal information, enabling the model to adapt to complex changes in targets and scenes. Experimental results on multiple datasets such as TrackingNet, GOT-10k, LaSOT and UAV123 show that the proposed method can track the target accurately and robustly, and the optimal results are obtained on GOT-10k dataset. AO, SR_0.5 and SR_0.75 were 73.7%, 83.8% and 70.6%, respectively.

Mixed attribute data is one of the most common types of datasets, and clustering algorithms tailored for this type of data are a research hotspot in clustering analysis. Due to the advantages of spectral clustering algorithms in handling clustering problems of arbitrary-shaped data and converging to global optimal solutions, an improved spectral clustering algorithm(improved Jaccard and Mahalanobis-spectral clustering, IJM-SC) from the perspective of similarity measurement formulas is proposed.Based on the ideas of Jaccard distance and Mahalanobis distance, a similarity measurement suitable for mixed attribute data is designed, and its application in spectral clustering of mixed attribute data is explored. The developed algorithm is applied to cluster three mixed attribute datasets including the UCI heart disease dataset, demonstrating its superiority in clustering mixed attribute data. By comparing the performance metrics with existing algorithms, the results demonstrate that the proposed algorithm achieves better clustering of mixed attribute data.

Multimodal sentiment analysis, an inherently challenging research area, integrates textual, audio, and visual modalities to analyze human emotional tendencies. Existing studies suggest that the textual modality plays a dominant role in sentiment prediction. However, this predominance raises a potential issue: during training, machine learning models tend to learn spurious correlations between the input and the output, leading to an overreliance on textual information. This overreliance may cause models to incorrectly model spurious correlations between textual and sentiment labels, thus undermining the model's generalization ability. To address this challenge, an innovative counterfactual text debiasing(CFTB) algorithm is proposed for multimodal sentiment analysis. Our framework first employs causal graph to thoroughly analyze the causal relationships among the three modalities and the sentiment labels.Then, an auxiliary textual model is designed to precisely quantify the direct impacts of the textual modality to sentiment prediction and leverage an attention mechanism to accurately capture textual features that might introduce spurious correlations. During the inference phase, the CFTB algorithm demonstrates its unique advantage: it intelligently isolates the negative impacts caused by spurious textual associations from the overall multimodal information, while retaining and enhancing the beneficial information within the textual modality that genuinely contributes to sentiment prediction. Experiments on the MOSEI and MOSI datasets show that this framework can be effectively integrated into existing methods and has good generalization performance.

When mobile robots perform path planning, the traditional or classical ant colony algorithm often encounters problems such as fewer movement directions, smaller fields of view, non-optimal paths, and unsmooth paths. Aiming at the inherent shortcomings of the ant colony algorithm mentioned above, a parallel bidirectional 24 neighborhoods 16 directions ant colony algorithm is proposed. First, the 24 neighborhoods 16 directions path search method can expand the field of view of path search. Second, combining the 24 neighborhoods 16 directions path search method with the bidirectional alternating search strategy can better reach the endpoint and enhance the global search ability. Subsequently, the heuristic function includes starting point, current point, candidate node, and endpoint, as well as adaptive factors. At the same time, an improved transition probability formula is introduced to enhance the guidance of path search. Then, the crossing strategy is introduced to avoid getting stuck in local optima. Finally, the path node transfer strategy is adopted to smooth the path, resulting in fewer inflection points and the shortest path. On grid maps with different complexity, the improved ant colony algorithm proposed in this paper was compared with the traditional ant colony algorithm and other improved ant colony algorithms through simulation experiments. The simulation results proved that the algorithm proposed in this paper is feasible and effective.

Aiming at the disadvantage that it is hard to obtain multiple Pareto sets in solving multimodal multi-objective optimization problems, a decomposition-based differential evolution algorithm is presented. In the proposed algorithm, multiple individuals that are assigned to the same weight vector form a subpopulation for finding multiple different Pareto sets. Then, an environmental selection method is designed to locate multiple different Pareto optimal solutions in the subpopulation. Finally, two differential evolution strategies are utilized to generate the offspring. The simulation results of the IEEE CEC 2019 benchmark test suite show that the proposed algorithm has good distribution ability in the decision space and can find more Pareto optimal solutions.

The traditional flower pollination algorithms tend to exhibit poor exploitation when facing some complex optimization problems. Aiming at this weakness of the traditional flower pollination algorithms, an adaptive guidance flower pollination algorithm (AGFPA) is proposed. In the proposed AGFPA, an adaptive guidance mechanism (AGM) is introduced, which combines the global best individual surrounding strategy and the global best approaching strategy.The introduced AGM adaptively utilizes the global best individual to guide the population evolution, enhancing the exploitation capabilities while preserving the population diversity as much as possible. Specifically, the global best individual surrounding strategy focuses on exploiting the neighborhoods around the global best individual. Meanwhile, the global best approaching strategy utilizes the global best individual to guide the search directions, enabling the algorithm to explore a wide unknown area. In addition, an adaptive parameter control strategy is presented in the proposed AGFPA. The two key parameters, global pollination transform probability and step size factor, are adjusted according to the needs of different evolution stages, maintaining a good balance between exploitation and exploration. To test the performance of AGFPA, 18 benchmark functions are utilized in the experiments, which are commonly used in the field of swarm intelligence. The effectiveness of the strategies is discussed. Moreover, AGFPA is compared with several existing flower pollination algorithms and particle swarm optimization algorithms. Additionally, AGFPA is also used to estimate the fermentation kinetic parameters in the biochemical engineering. The experimental results show that AGFPA can exhibit promising performance on the most unimodal, multimodal and complex functions. Moreover, AGFPA can yield excellent results in the biochemical engineering applications.