Graph-based Retrieval-Augmented Generation (GraphRAG) is a prominent technique for mitigating hallucinations in Large Language Models (LLMs), yet it faces critical limitations in high-stakes anesthesiology question answering. At the retrieval stage, pervasive synonymy and knowledge fragmentation cause clinical queries to resolve to incorrect synonym clusters, while the highly specialized and domain-specific nature of clinical language widens the semantic gap, making an effective balance between recall and precision difficult to achieve. At the reasoning stage, the high structural heterogeneity of anesthesiology knowledge graphs, combined with the strong context-dependence of clinical semantics, renders existing multi-hop reasoning susceptible to clinical logic drift along weak associations, thereby introducing unreliable reasoning paths. Finally, at the fusion stage, conflicts between retrieved evidence and the model's parametric knowledge can easily lead to highly confident yet clinically unsafe hallucinations. To address these challenges, we propose the Description-augmented, Uncertainty-aware, and Adaptive Layered Knowledge Framework (DUAL-Know). For retrieval, query augmentation and description-augmented semantic recall bridge terminology gaps and comprehensively retrieve relevant knowledge. For reasoning, a Dynamically Gated Heterogeneous Multi-Head Attention (DGHMA) mechanism accommodates graph heterogeneity and dynamically injects query intent into each reasoning layer, followed by an uncertainty-aware path ranking module that filters noisy chains. For fusion, a multi-metric scoring strategy jointly evaluates model confidence, retrieval consistency, and semantic coherence to safely arbitrate knowledge conflicts. Experiments demonstrate that DUAL-Know consistently outperforms strong baselines, generating more accurate, reliable, and verifiable answers for complex clinical question-answering tasks.

Brain tumor localization and segmentation from MRI is a challenging task in the field of Medical AI. With recent advancements, various techniques have been developed to assist medical professionals in detecting brain tumors using Artificial Intelligence. Although machine learning algorithms have shown efficiency in tumor segmentation, they often lack interpretability, making it difficult to trust and validate their predictions. In this paper, we developed an interpre table UNet model for brain tumor segmentation, incorporating Gradient-weighted Class Activation Mapping (Grad-CAM) and Shapley Additive Explanations (SHAP) to enhance model transparency. The BraTS2020 benchmark dataset was used for training and evaluation. The model demonstrated strong performance, achieving an accuracy of 99.35%. Grad-CAM was applied to visualize the regions of interest identified by the model, enhancing interpretability. Additionally, the SHAP library was utilized to explain the predictions of multiple machine learning models (Random Forest, KNN, SVC, MLP) employed for estimating patient survival days, further contributing to the model’s transparency and clinical utility.

Chinese medical named entity recognition aims to identify named entities from unstructured Chinese medical texts, which has received increasing attention from both academia and industry. The existing efforts could be divided into three categories: rule-based, machine learning-based, and deep learning-based approaches. However, there are still two limitations in Chinese medical named entity recognition: insufficient information fusion and ambiguous word segmentation, leading to incomprehensive representations and suboptimal performance. To tackle these problems, we propose a novel model—Semantic Enhanced relational graph attention Network for Chinese medical named Entity recognition (SENCE), which considers morphology-level, character-level, and word-level semantics simultaneously. Specifically, we build semantic graphs to encode semantic information on different levels and develop semantic enhanced relational graph attention network to capture high-order information and enhance relation-aware representations. Finally, we feed the enhanced representations into bidirectional long short-term memory (BiLSTM) and conditional random field (CRF) for named entity label prediction. We conducted extensive experiments on two benchmark datasets. The results illustrate that the proposed model outperforms state-of-the-art methods, demonstrating the effectiveness of the proposed model. Further ablation studies validate the rationality of the components.

Delirium is an acute and reversible neuropsychiatric syndrome, with its diagnosis relying on clinicians’ dynamic assessments of patients. This process requires the integration of multidimensional clinical information, for which electronic medical records (EMRs) serve as a crucial source. However, Chinese EMRs are often lengthy, unevenly structured, and exhibit considerable variability in symptom descriptions, leading to key diagnostic information being obscured by redundant narratives. These characteristics pose significant challenges for traditional methods to consistently extract relevant diagnostic features. This paper proposes a deep learning framework, guided by medical knowledge, to improve the classification performance and interpretability of delirium in Chinese long-form EMRs. The framework employs a region-guided chunking mechanism to segment records into semantically coherent units and extract key features. It combines differentiated data augmentation and contrastive learning methods to strengthen semantic discriminability, and utilizes adaptive feature fusion to achieve cross-chunk collaborative decision-making. Experimental results demonstrate that this method effectively enhances the accuracy and reliability of delirium diagnosis in Chinese medical texts, showing strong clinical applicability.

Diabetic retinopathy (DR) is the leading cause of blindness in diabetic patients, which necessitates precise grading of retinal lesions for early diagnosis. Existing DR grading methods typically employ image enhancement techniques to improve the quality of fundus images. However, due to variations in imaging devices and differences in the proficiency of medical practitioners, the quality of images often exhibits significant heterogeneity. Uniform enhancement across all fundus images may inadvertently amplify noise artifacts, particularly in high-quality images. Moreover, since diabetic lesions in fundus images are often small, reliance solely on global image features makes it difficult to fully capture fine-grained lesion features. To address these challenges, this paper proposes a dual-stream deep learning model that integrates quality-aware dynamic enhancement and a multi-instance multi-scale vision transformer. First, An image quality assessment-based selective enhancement strategy was implemented, wherein only low-quality fundus images underwent enhancement processing. Then, a dual-branch processing architecture is designed to differentially handle enhanced and non-enhanced images. Experimental results on real-world datasets demonstrate the effectiveness of the proposed method.

In the medical treatment of pneumonia, most patients undergo multiple CT scans at different time points, generating serialised CT data. However, traditional segmentation methods based on convolutional networks are difficult to capture inter-slice dependencies, which can result in missing sequence information. Additionally, Transformer-based segmentation methods tend to overemphasize temporal information, leading to the loss of local features in a single image. To address these issues, we propose a Serialised Pulmonary Lesion Segmentation with Spatiotemporal Transformer (SPLS-STT). Specifically, we firstly capture local features of lesion regions and global features from the serialised lung CT through a self-attention mechanism, which generates a spatial feature maps. Secondly, the obtained spatial feature maps at multiple time points are arranged in chronological order, and the dependencies between the pneumonia lesions at different time points are established by the cross-window temporal attention mechanism. Finally, we combine the features in the temporal and spatial dimensions through the cross-attention mechanism to compensate for the loss of local features during feature extraction in the temporal dimension. Experimental results show that our spatiotemporal Transformer model outperforms the traditional single-dimensional processing methods in terms of accuracy metrics. Furthermore, our model achieves a Dice coefficient of 37.9% on the lung CT sequence dataset, which is an improvement of 9.5% compared to the Vision Transformer.

Wei Zhai, Nan Bai, Qing Zhao, Jianqiang Li, Fan Wang, Hongzhi Qi, Meng Jiang, Xiaoqin Wang, Bing Xiang Yang, Guanghui Fu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Accurate prediction of intraoperative hazardous events and generation of effective intervention plans are critical to surgical safety, but face multiple challenges of real-time, accuracy, and interpretability. Large-scale language models have potential, but their high cost and potential ‘illusion’ problems limit their application in real-time clinical environments. Traditional multitask learning models are efficient but knowledge-constrained, making it difficult to capture complex reasoning processes. To bridge this gap, this paper proposes a multi-objective distillation knowledge enhancement model-KEMO, which innovatively adopts a multi-objective chain-of-thought distillation framework to not only mimic the prediction results of the instructor’s LLM, but also explicitly migrate its structured reasoning process to the lightweight student model, which improves the answerability of the model by synergistically optimising the three objectives of event prediction, reasoning alignment and scenario generation. Interpretability. Meanwhile, combined with the Knowledge Graph-based Retrieval Augmented Generation mechanism, validated medical knowledge is dynamically injected to enhance the accuracy and reliability of decision-making and reduce model illusion. The experimental results show that the KEMO model significantly outperforms traditional models of the same magnitude in intraoperative hazardous event prediction and prognostic proposal generation, and achieves a performance comparable to that of a large faculty model.The KEMO model effectively bridges the gap between the large language model and the actual clinical application, and facilitates the transformation of the large model knowledge to the actual clinical deployment.

Accurately determining an individual's preference for images remains a major challenge in the field of emotional research. This study proposes a novel paradigm for identifying individual image preferences using electroencephalography (EEG) signals and brain-computer interface (BCI). The paradigm involves both explicit and implicit tasks, where participants perform a typical event-related potential-based brain-computer interface(ERP-BCI) operation and their subjective image preferences are identified, respectively. Two experiments with a total of 27 participants demonstrate that event-related potential (ERP) signals during explicit BCI tasks are significantly influenced by target image preferences, enabling high-accuracy image preference recognition. Online experiments selecting positive and negative preference images from a candidate pool show top-1 accuracy approaching 100% and top-3 accuracy exceeding 90%. These results indicate the effectiveness of the proposed EEG-based image preference recognition paradigm, laying the groundwork for preference analysis applications.

Depression is a prominent public mental health issue affecting millions globally. With the growth of online communities, many individuals suffering from depression are increasingly seeking guidance and support from community-based Question-Answering (QA) systems. Previous methods, which recommend the best answer by calculating the semantic distance between a question and its potential answers, seldom incorporate psychological indicators as features. Answers from non-professionals might offer misguided information to those with depression. In our approach, we employ a multi-task learning method, taking into account both semantic information and psychological indicators for answer selection in depression community QA systems. To compute the semantic similarity between a given question and its potential answers, we gather surface-layer textual information. We then evaluate these candidate answers using psychological indicators from two dimensions: (1) employing five expert-defined comprehensive quality indicators to assess answer quality, and (2) using existing standard answers, provided by experts for various question categories, as external references to automatically review the candidate answers. Finally, by integrating the semantic similarity and psychological indicators, we make the final answer selection. Experimental results have shown that our model significantly enhances answer recommendations in the mental health domain.

In the social media, users frequently express personal emotions, a subset of which may indicate potential suicidal tendencies. The implicit and varied forms of expression in internet language complicate accurate and rapid identification of suicidal intent on social media, thus creating challenges for timely intervention efforts. The development of deep learning models for suicide risk detection is a promising solution, but there is a notable lack of relevant datasets, especially in the Chinese context. To address this gap, this study presents a Chinese social media dataset designed for fine-grained suicide risk classification, focusing on indicators such as expressions of suicide intent, methods of suicide, and urgency of timing. Seven pre-trained models were evaluated in two tasks: high and low suicide risk, and fine-grained suicide risk classification on a level of 0 to 10. In our experiments, deep learning models show good performance in distinguishing between high and low suicide risk, with the best model achieving an F1 score of 88.39%. However, the results for fine-grained suicide risk classification were still unsatisfactory, with the best weighted F1 score of 50.89%. To address the issues of data imbalance and limited dataset size, we investigated both traditional and advanced, large language model based data augmentation techniques, demonstrating that data augmentation can enhance this model performance by up to 4.65% points in F1-score. Notably, the Chinese MentalBERT model, which was pre-trained on psychological domain data, shows superior performance in both tasks. This study provides valuable insights for automatic identification of suicidal individuals, facilitating timely psychological intervention on social media platforms. The source code and data are publicly available at: https://github.com/HongzhiQ/FineGrainedSuicideDetection.

OBJECTIVE: The auditory event-related potential based brain-computer interface (aERP-BCI) is a classical paradigm of brain-computer communication. To improve the coding efficiency of aERP-BCI, this study proposes a method using two parallel voice channels to add the coding dimension based on the cocktail party effect. METHODS: The novel paradigm used male and female voices to establish two parallel oddball sound stimulus sequences. In comparison, the baseline paradigm only presented male or female stimulus sequences. Both the double voice condition (DVC) and the single voice condition (SVC) paradigms carried out offline experiments and the DVC also carried out online experiment. Subsequently, the EEG signal and BCI operation results were compared and analyzed. CONCLUSION: The cocktail party effect caused a significant difference in the EEG responses of non-target stimulus between the focused vocal channel and the ignored vocal channel under the DVC paradigm, and the focused and ignored channels achieved a recognition accuracy of 97.2%. The target recognition rate of DVC was 82.3%, with no significant difference compared with 85% of SVC while the information transfer rate (ITR) of DVC reaching 15.3 bits/min was significantly higher than that of SVC. SIGNIFICANCE: The cocktail party effect improves the coding efficiency by adding parallel channels without reducing the target/non-target stimulus recognition in the focused vocal channel. This provides a novel direction for the performance improvement of aERP-BCI.

In recent years, the brain-computer interface (BCI) based on motor imagery (MI) has been considered as a potential post-stroke rehabilitation technology. However, the recognition of MI relies on the event-related desynchronization (ERD) feature, which has poor task specificity. Further, there is the problem of false triggering (irrelevant mental activities recognized as the MI of the target limb). In this paper, we discuss the feasibility of reducing the false triggering rate using a novel paradigm, in which the steady-state somatosensory evoked potential (SSSEP) is combined with the MI (MI-SSSEP). Data from the target (right hand MI) and nontarget task (rest) were used to establish the recognition model, and three kinds of interference tasks were used to test the false triggering performance. In the MI-SSSEP paradigm, ERD and SSSEP features modulated by MI could be used for recognition, while in the MI paradigm, only ERD features could be used. The results showed that the false triggering rate of interference tasks with SSSEP features was reduced to 29.3%, which was far lower than the 55.5% seen under the MI paradigm with ERD features. Moreover, in the MI-SSSEP paradigm, the recognition rate of the target and nontarget task was also significantly improved. Further analysis showed that the specificity of SSSEP was significantly higher than that of ERD (p < 0.05), but the sensitivity was not significantly different. These results indicated that SSSEP modulated by MI could more specifically decode the target task MI, and thereby may have potential in achieving more accurate rehabilitation training.

Although the inter-subject information has been demonstrated to be effective for a rapid calibration of the P300-based brain-computer interface (BCI), it has never been comprehensively tested to find if the incorporation of heterogeneous data could enhance the accuracy. This study aims to improve the subject-specific P300 classifier by adding other subject's data. A classifier calibration strategy, weighted ensemble learning generic information (WELGI), was developed, in which elementary classifiers were constructed by using both the intra- and inter-subject information and then integrated into a strong classifier with a weight assessment. 55 subjects were recruited to spell 20 characters offline using the conventional P300-based BCI, i.e. the P300-speller. Four different metrics, the P300 accuracy and precision, the round accuracy, and the character accuracy, were performed for a comprehensive investigation. The results revealed that the classifier constructed on the training dataset in combination with adding other subject's data was significantly superior to that without the inter-subject information. Therefore, the WELGI is an effective classifier calibration strategy which uses the inter-subject information to improve the accuracy of subject-specific P300 classifiers, and could also be applied to other BCI paradigms.

The use of hand gestures provides an attractive alternative to cumbersome interface devices for human-computer interaction (HCI). However, the number of hand gestures has not been fully explored for HCI application. It is necessary to achieve more gestures as the command of interface. This paper proposed a method to recognize nine different hand gestures. Using camera to get images of people wearing pink gloves, and then preprocess those images by color splitting, morphological processing and edge extraction. Fourier descriptor, edge histogram and boundary moment invariants are three methods of feature extraction. At last, the template matching was used to realize the hand gesture recognition. The average recognition rate of the nine different gestures employing three different methods is 0.859.