Personalized drug recommendation is a process that analyzes the individual characteristics of patients to predict their most suitable drug plan. However, existing drug recommendation methods use the information fusion method of vector horizontal concatenation, which is computationally complex and introduces unnecessary noise. Furthermore, previous models only rely on the relationship between drugs and symptoms to achieve the recommendation process from symptoms to drugs, ignoring the semantic interaction information between diseases and symptoms. In addition, existing methods cannot clearly determine which drugs should be retained or removed when faced with drug interaction conflicts, it is difficult for them to reduce drug-drug interactions (DDI) for recommendation. To solve these problems, we propose a personalized drug recommendation model named DFDC-GNN. In the model, we design a dimensionality reduction processing module to achieve a semantic information fusion of a lower dimension and obtain a comprehensive embedding representation for drug recommendation. Furthermore, we introduce a disease representation module and use a graph attention network to capture high-order semantic relationships among the nodes of symptoms, disease, and drugs. Moreover, we design a drug constraint decision mechanism to effectively extract the interaction relationships between drugs and reasonably control DDI in the drug combination. Finally, we conducted experiments on the MIMIC-III dataset to compare with other baseline models.

The function-level code generation is an important task in the combination of software engineering and artificial intelligence, which aims to improve the productivity of software development by automatically generating function-level code based on task descriptions. However, this task currently suffers from several problems: 1) in the fine-tuning phase, current large language models cannot sufficiently capture the detailed syntactic structure of the code dataset; 2) previous retrieval-augmented methods cannot adequately consider the complex dependencies between code snippets in external code repositories; 3) existing large language models introducing self-repair mechanism tend to overfocus on previously generated erroneous code during the self-repair process. To solve these problems, we propose a retrieval filtering and thought enhancement framework for function-level code generation based on large language model. In our model, we design an abstract syntax tree mapping preprocessing module to preprocess the dataset and help large language models learn the detailed syntax information of the dataset. Furthermore, we design a retrieval filtering and thought enhancement module to retrieve the most relevant snippets of code for the task and to enhance the chain-of-thought of our model. In addition, we design a self-repair mechanism to prevent large language models from overfocusing on generated erroneous code, helping them explore more solutions to repair the erroneous code. We experimented and evaluated our model on HumanEval, MBPP, and MultiPLE benchmarks to compare with other baseline models.

The generation of radiation reports is an essential task in the field of medical artificial intelligence which aims to automatically generate text descriptions of radiology images. However, there are still several problems in this task: 1) existing methods lack global feature interaction when extracting image features, and their ability to represent images is limited; 2) previous models need to retrieve similar triplets input models from pre-constructed knowledge graphs, and the triplets lack entity relationship refinement; 3) existing approaches lack a correlation mechanism across samples that makes it difficult to effectively capture small abnormal regions; 4) the self-attention mechanism of the Transformer decoder is good at capturing global dependencies but ignores relationships between local contexts. To address these issues, we propose a triplet optimization and difference detail perception network with adaptive feature enhancement. In our model, we design an adaptive image feature enhancement module to dynamically capture global image features. Furthermore, we propose a multi-modal triplet optimization module that boosts capability for detecting abnormal regions by incorporating context-aware entity relationship refinement into the initial triplet. Moreover, we design a difference comparison weighting module to obtain fine-grained features between different samples and improve cross-sample correlation so that the model pays more attention to small details and anomalies that are easy to ignore. Finally, we design a detail-aware enhancement decoder to make the decoder pay more attention to the relationship between local contexts. We experimented and evaluated our model on the IU-Xray and MIMIC-CXR datasets to compare with other baseline models.

Document-level entity relation extraction is an important task in the field of natural language processing, which plays an important role in semantic understanding and knowledge graph construction. However, existing deep neural networks and graph neural networks models are limited by their performance and parameters number, which can not capture global semantics and have poor generalization ability. Furthermore, existing methods employing large language model for entity relation extraction do not establish good relationships among multi-tasks of entity relation extraction task, resulting in more information can not be effectively shared and transmitted between tasks. In addition, previous approaches can not effectively eliminate false entities and relationships. To solve these problems, we propose an entity relation extraction framework based on large language model and multi-tasks iterative prompt engineering. In our model, we design an iterative prompt engineering, which can better establish the relationship among multi-tasks, and ensure every task to obtain the optimal results. Moreover, we design semantic merging, group disambiguation and self-verification modules to eliminate the false entity relations and noise nodes. Additionally, we design summary prompts to provide sufficient global semantics for better text segmentation. Finally, we evaluated our model on wikiann, wikineural, ACE2005, CoNLL2003, CoNLL2004, and SciERC datasets and compared it with other baseline models.

3D medical image segmentation is an essential task in the medical image field, which aims to segment organs or tumours into different labels. A number of issues exist with the current 3D medical image segmentation task: existing models cannot simultaneously obtain the space correlation and depth correlation of 3D slices; previous models suffer from local detail loss of positional embedding in 3D images; previous approaches often have blurring of boundaries in segmenting 3D images. To solve these shortcomings, we propose a 3D medical image segmentation model named TPM-SAM. In our model, we design a twinchannel image encoder to simultaneously capture the space correlation and depth correlation of 3D slices through a multi-head attention mechanism and improved adapters. Furthermore, we design a prompt encoding generator, which divides the volumetric image into small blocks and better captures the local detail information. In addition, we introduce a multi-layer aggregation decoder by employing U-Net with multi-level skip connection to solve the blurring of boundaries in processing 3D images. Finally, we experimented and evaluated our model on KiTS21 and LiTS17 datasets to compare with other baseline models.

Radiology report generation is an essential task in the medical field, which aims to automate the generation of medical terminology descriptions of radiology images. However, this task currently suffers from several problems: 1) existing methods need to manually build knowledge graphs or templates (consuming time and effort) to introduce medical or prior knowledge to assist in report generation; 2) previous models cannot handle the problem of data bias well (anomaly reports and anomaly descriptions make up only a tiny portion of the dataset), causing the models to ignore the learning of anomaly descriptions easily; 3) existing approaches cannot robustly supervise the model, resulting in incomplete and inconsistent reports being generated. To address these issues, we propose a cross-modal interactive memory network based on fine-grained medical feature extraction. In our model, we design a cross-modal interactive memory network to automatically store and remember the required medical text knowledge and use this medical knowledge to help generate reports. Furthermore, we design an abnormal medical knowledge enhancement module to enhance the learning of abnormal fine-grained knowledge through the interaction of disease topics and their states to interact with text features. In addition, we design a cross-modal joint semantic loss unit to reduce semantic differences between different features and improve the visual representation ability of the model. We experimented and evaluated our model on MIMIC-CXR and IU-Xray datasets to compare with other baseline models.

Personalized news recommendation is the process of predicting the relevance of news to users and recommending news to user to fulfill their information needs. However, existing news recommendation methods extract semantic information from users and candidate news respectively, ignoring semantic interaction information between users and candidate news. Furthermore, previous models only use same node types for message passing, ignoring different characteristics and topology between different node types. In addition, existing methods learn news representations through text representations, ignoring semantic correlation information between entity relationships and texts. To solve these problems, we propose a personalized news recommendation model named CoHG. In our model, we design a collaborative fusion module to obtain semantic interaction information through interacting user history news with candidate news. Furthermore, we design a heterogeneous gated graph neural network that maps different node types into a same space to extract higher-order information in user graphs for message passing. Moreover, we design an enhanced relevant attention module to enhance semantic correlation information of text content by aggregating text representation and entity representation into a unified representation. Finally, we conducted experiments on MIND and Adressa datasets to compare with other baseline models.

The fine-grained image classification task is a major task in computer vision. Although many deep learning inter-pretable models have been proposed for this task, the accuracy and interpretability of these models need to be improved. We propose an interpretable fine-grained image classification model based on an improved neural prototype tree. In our model, we design the new multi-grained feature extraction network with three new backbone networks to extract features of fine-grained and multi-grained images more effectively. Furthermore, we design a new background prototype removing mechanism in the soft neural binary decision tree layer to optimize prototype path decision. Afterwards, we design a new loss function with both a leaf node loss function and a fully connected layer loss function to improve the generalization ability. Finally, we evaluate our model on three public datasets CUB-200-2011, FGVC-Aircraft, and Chest X-ray to compare with other baseline models.

Commonsense question answering (CSQA) is a challenging learning task that aims to give correct answers to commonsense questions. CSQA models combining large pretrained language models with knowledge graphs are proposed to perform one-way or two-way information fusion to enhance their commonsense reasoning ability. However, existing CSQA models only fuse local information at the word level, ignoring the global semantic information fusion. Furthermore, current CSQA models often introduce noise nodes when constructing the knowledge subgraph. In addition, existing methods neglect the edge information in message aggregation. To solve these shortcomings, we propose a novel CSQA model named MDEQA. In our model, we design the multi-layer attention fusion module to bidirectionally fuse the word-level local information and global semantic information of question context and knowledge subgraph. Moreover, we design the dynamic graph neural network module with improved GAT and aggregating edge information to form the dynamic subgraphs which alleviate the interference of noise nodes on reasoning and enhance the commonsense reasoning ability of our model. Finally, we evaluated our model on CommonsenseQA and OpenBookQA datasets to compare with other baseline models.

Medical image classification is a main task in medical diagnosis field. Some black box models have achieved expert-level accuracy on medical datasets, but these models are less adopted in clinical practice due to the lack of interpretability. In the past few years, designing interpretable models has been one of the major challenges in the medical field. The existing interpretable prototype networks only use medical images for training, ignoring the role of medical reports, and these medical reports can assist in prototype training and activation. Furthermore, existing prototype network methods neglect the position information in medical images, which is helpful for disease diagnosis. To solve these shortcomings, we propose an interpretable medical image classification framework (MProtoNet) that improves the accuracy and interpretability of disease predictions. In MProtoNet, we design a multimodal attention module and use prototype activation restriction loss to provide evidence for prototype training and activation. In addition, we design a position embedding module and multi-factor similarity calculation method to effectively utilize the position information in the image. We conducted experiments on the chest datasets MIMIC-CXR and open-I to test the model and compare it with other baseline models. Experimental results show that MProtoNet has made improvements in accuracy while preserving the interpretability of the model.

The significant success of deep learning has sparked interest in its application in medical diagnosis. Some deep learning models have achieved expert-level accuracy on some medical datasets, but these models are rarely used in clinical practice due to the lack of interpretability. Therefore, the research topic of explainable artificial intelligence (XAI) has emerged to make the reasoning process of the model transparent and interpretable. In this case, we applied interpretable artificial intelligence to dermatoscopy image diagnosis for the first time. Specifically, we use an interpretable prototype network for dermoscopy image diagnosis. To solve the problem of weak generalization performance of a single network, we propose to construct a new prototype network using the dual-path network. Besides, we propose a new gate similarity calculation method to reduce the activation of low-similarity regions, thereby reducing the generation of inaccurate prototypes and improving the diagnostic ability of the model. We conducted experiments on the dermoscopy datasets HAM10000 to test the model and compare it with other baseline models. Experimental results show that DP-ProtoNet has made improvements in accuracy while preserving the interpretability of the model.

Biomedical commonsense question answering is a challenging learning task that aims to give correct answers to biomedical commonsense questions. Many works have used rule-based or deep learning approachs to accomplish this task. Recently, an extensive research path is that pre-trained language models combined with graph neural networks (GNNs) to improve the accuracy of biomedical commonsense question answering. However, GNN is prone to the over-smoothing problem, causing the models to lose the ability to reason. In order to alleviate the over-smoothing problem and improve the inference ability for biomedical commonsense question answering, we propose a new end-to-end model named BiomGIN. In BiomGIN, we introduce the Graph Optimal Transport Networks (GOTNet) to use node-centroid attention to capture non-local messages in the knowledge graph, which alleviates the model over-smoothing problem. In addition, we design a question parsing module based on Transformer to generate linguistic instructions, which enhances the inference capability of the GNN. Finally, we evaluated our model on MedQA-USMLE dataset to compare with other baseline models. The experimental results demonstrate the method proposed in this paper achieves state-of-the-art results.

Transcriptome prediction from genetic variation data is an important task in the privacy-preserving and biometrics field, which can better protect genomic data and achieve biometric recognition through transcriptome. Many transcriptome prediction methods have achieved good accuracy from genetic variation data. However, these traditional transcriptome prediction methods have the problems of linear assumption, overfitting, expose personal privacy, and extensive manual optimization. To solve these shortcomings, we propose an attention-based transcriptome prediction model from genetic variation named RATPM that improves the accuracy of transcriptome prediction and protects participant genomic data. In RATPM, we introduce and improve the deep learning model with multi-head self-attention into the transcriptome prediction stage of Predixcan, which uncovers the non-linear relationship between genetic variation and transcriptome. Moreover, we introduce a residual attention module to generate attention-aware features and extract more accurate features at different levels from genetic variation. Furthermore, we introduce the BERT pre-training module to encode genetic variation fully utilizing their contextual information. Our research enables scientific institutions to publish only predicted transcriptomic data for biometric purposes, thus protecting the genomic information of the subjects. Finally, we evaluated our model on the 1000 Genomes and Geuvadis projects datasets to compare with other baseline models.

Link prediction of knowledge graph is a relatively widely studied task in knowledge graph completion, the purpose of which is to complete the incomplete triples according to the original knowledge triples of the knowledge graph. To solve the problem of handling the heterogeneous neighborhood and the injective problem in the link prediction of knowledge graph, we propose a novel deep learning model called Transformation Assumptions with Message Passing Aggregation Network (TMPAN). TMPAN can effectively deal with the heterogeneous neighborhood information by introducing TransGCN’s transformation assumptions into DPMPN, which transforms head entities to tail entities using relationships as transformation operators. TMPAN also solves the injective problem caused by the single-aggregation operation by employing the multiple aggregators of the Principal Neighborhood Aggregation network (PNA) model. We comprehensively evaluate our model compared to typical baseline models by conducting experiments on two public datasets, FB15K-237 and YAGO3-10. The experimental results show the effectiveness of our model in the link prediction task.

Biomedical event extraction is a fundamental information extraction task, which aims to identify biomedical event triggers and parameters in the text. Although various deep learning and Graph Convolutional Network (GCN) models have been proposed for this task, these models are insufficient to acquire enough local and global context information of documents. To effectively extract joint local and global context information, we propose a joint biomedical event extraction model named BGHGCN, which consists of Bi-directional Long Short-Term Memory (BiLSTM), improved BiAffine Graph Parser (IBGP), GCN and hypergraph convolutional networks (HGCN). Our model employs BiLSTM to learn word sequence features and uses improved BiAffine Graph Parser to enrich dependent syntax features. Afterwards we use GCN to extract local features from IBGP and BiLSTM. Specifically, we introduce HGCN to jointly extract local and global context information with a new fusion mechanism of local feature and incidence matrix, which can effectively extract structural features of hypergraph including node and hyperedge features. Finally, we evaluated our model on two biomedical event datasets MLEE and GE to compare with other baseline models.

Joint 2D object detection and 3D reconstruction is an essential computer vision task to get more accurate detection and representation model of the target object. We proposed a novel joint 2D object detection and 3D reconstruction model that enhances the ability of the 2D object detection and the 3D reconstruction, called Adversarial Fusion Mesh Region Convolutional Neural Networks (AFM R-CNN). Our proposed model introduces the Deep Convolutional Generative Adversarial Network (DCGAN) to generate adversarial images and input the real and adversarial images into the object detection module GA-RPN to determine the position and anchor box of the target object. Next, to make better use of the two-dimensional information of the image, the voxel conversion and Fusion model Pix2Vox is introduced to fuse the two types of image features and generate coarse voxels. Afterwards, to differentiate the voxel information more efficiently, we use the Principal Neighborhood Aggregation network (PNA) model in 3D model refinement. The contrast experimental results on the open domain dataset (Pix3D) with baseline models demonstrate the effectiveness of AFM R- CNN in joint 2D object detection and 3D reconstruction task.