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.

Cushing’s syndrome is caused by excessive glucocor-ticoid secretion and often presents with facial features such as moon facies and plethora, making facial images valuable for diagnosis. Recent studies have used pre-trained CNNs for automated diagnosis using frontal facial images. However, CNNs focus on local features and may miss global facial characteristics typical of Cushing’s syndrome. Transformer-based visual models, through self-attention mechanisms, are better suited for capturing global features. The foundational model DINOv2, also based on the vision Transformer architecture, has recently attracted attention. In this study, we compared the performance of various pre-trained models, including CNNs, Transformer-based models, and foundational models like DINOv2, in a transfer learning setting. We also investigated biological sex bias and the effect of freezing mechanisms on DINOv2. Our results show that Transformer-based models, especially ViT, and DINOv2 outperformed CNNs, with ViT achieving the highest F1 score of 85.74%. All models showed higher accuracy for female samples, likely due to biological sex imbalance. Freezing mechanisms notably improved DINOv2 performance. In conclusion, Transformer-based models and DINOv2 show strong potential for classifying Cushing’s syndrome using facial images.The source code is publicly available at: https://github.com/songchangwei/Cushing-Disease-Diagnosis.

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.

Mental health is a significant global public health issue, and psychological support hotlines play a crucial role in providing mental health assistance and identifying suicide risks at an early stage. However, the emotional expressions conveyed during these calls remain underexplored in current research. This study introduces a novel method that combines pitch acoustic features with deep learning-based features to analyze and understand emotions expressed during hotline interactions. Using data from China's largest psychological support hotline, which includes 105 subjects, our method achieved an F1-score of 79.13% for negative binary emotion classification. Additionally, the proposed approach was validated on an open dataset for multi-class emotion classification, where it demonstrated better performance compared to the state-of-the-art methods. To explore its clinical relevance, we applied the model to analysis the frequency of negative emotions and the rate of emotional change in the conversation, comparing 46 subjects with suicidal behavior to those without. While the suicidal group exhibited more frequent emotional changes than the non-suicidal group, the difference was not statistically significant. Importantly, our findings suggest that emotional fluctuation intensity and frequency could serve as novel features for psychological assessment scales and suicide risk prediction. The proposed method provides valuable insights into emotional dynamics and has the potential to advance early intervention and improve suicide prevention strategies through integration with clinical tools and assessments. The source code is publicly available at: https://github.com/Sco-field/Speechemotionrecognition/tree/main.

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.

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.

Feature sparse problem is commonly existing in patient similarity calculation task with clinical data, to track which, some approaches have been proposed to use Graph Neural Network (GNN) to model the complex structural information in patient Electronic Medical Records (EMRs). These GNN based approaches usually treat medical concepts (i.e., symptoms, diseases) as nodes to learn spatial features and adopt Recurrent Neural Network (RNN) to learn temporal sequence of these concepts. However, in many cases, several sequential concepts contained in EMR text are considered as occur simultaneously in the clinical diagnosis (i.e., some symptoms are detected simultaneously by once test), learning temporal sequence of these sequential concepts might cause noise for patient similarity calculation. Furthermore, the limited discriminative capability of concepts cannot provide sufficient indicative information for similarity learning. To this end, we propose a Knowledge-guided Event-relation Graph Learning Network (KEGLN) for patient similarity calculation. Specifically, after event extraction, we first construct element-relation graphs and use the first Graph Convolutional Network (GCN) and Graph Attention Network (GAT) layer to aggregate features from each event and its involved elements for reducing the noise produced by temporal sequence of concepts. Meanwhile, the entity description and attribute-value structure are extracted to supplement background knowledge of elements (concepts and trigger words). For the updated event nodes, we then design a event-relation graph and adopt the second GCN and GAT layer to aggregate information from events and their directly neighbors to extract spatial features of events at the current moment. Finally, the Bidirectional Long Short-Term Memory (BiLSTM) model is adopted to learn temporal dependency of event nodes to capture the dynamic change of disease progress. Through diverse datasets and extensive experiments, our KEGLN model outperforms all baselines for Chinese patient similarity calculation.

Medical Named Entity Recognition is essential for structuring medical text data, thereby aiding in the creation of medical applications like knowledge graphs and diagnostic systems. Contemporary methods predominantly utilize models of word embedding, subsequently augmented by various models for semantic comprehension, to enhance entity recognition performance. However, in the medical domain, specialized terminologies pose a challenge for general domain word embedding models. Furthermore, current methods frequently neglect local semantic attributes and face challenges in fully grasping global semantic features, attributed to the dimensional constraints of word embeddings. To address these challenges, we propose the Stacked Attention Network (SAN) for Chinese medical NER. We fine-tune RoBERTa using real-world electronic medical record data to incorporate medical term features and utilize a CNN model to extract local semantic features. Furthermore, we introduce a stacked BiLSTM with a multi-layer structure to effectively capture global semantic information. Experimental results on real-world medical text data demonstrate that our SAN model achieves an F1 score of 91.5%, outperforming other advanced NER models.

Medical image analysis frequently encounters data scarcity challenges. Transfer learning has been effective in addressing this issue while conserving computational resources. The recent advent of foundational models like the DINOv2, which uses the vision transformer architecture, has opened new opportunities in the field and gathered significant interest. However, DINOv2's performance on clinical data still needs to be verified. In this paper, we performed a glioma grading task using three clinical modalities of brain MRI data. We compared the performance of various pre-trained deep learning models, including those based on ImageNet and DINOv2, in a transfer learning context. Our focus was on understanding the impact of the freezing mechanism on performance. We also validated our findings on three other types of public datasets: chest radiography, fundus radiography, and dermoscopy. Our findings indicate that in our clinical dataset, DINOv2's performance was not as strong as ImageNet-based pre-trained models, whereas in public datasets, DINOv2 generally outperformed other models, especially when using the frozen mechanism. Similar performance was observed with various sizes of DINOv2 models across different tasks. In summary, DINOv2 is viable for medical image classification tasks, particularly with data resembling natural images. However, its effectiveness may vary with data that significantly differs from natural images such as MRI. In addition, employing smaller versions of the model can be adequate for medical task, offering resource-saving benefits. Our codes are available at https://github.com/GuanghuiFU/medical_dino_eval.

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.

Breast cancer has become the primary cancer endangering women's life and health. A large number of evidence based on medicine show that early screening can effectively reduce the mortality rate of breast cancer. With the development of computer vision technology, the computer-aided diagnosis system for breast cancer screening and detection has attracted extensive attention from all walks of life and breast lesion localization and diagnosis are the key steps. Therefore, this article starts from the perspective of model feature construction to conduct an in-depth analysis and a comprehensive summary of the existing research on breast image lesion localization and benign-malignant diagnosis. This article divides feature construction methods into three types based on novel perspectives: domain knowledge-driven, data-driven, and domain knowledge-driven fusion data-driven. On this basis, the article uses a systematic review method to classify, summarize, and compare the models for breast image lesion localization and diagnosis which greatly expands and deeply analyzes existing reviews. In addition, this article elaborates on the existing problems in current research work and discusses the future outlook of breast image lesion localization and benign-malignant diagnosis.

Suicide and suicidal behaviors remain significant challenges for public policy and healthcare. In response, psy-chological support hotlines have been established worldwide to provide immediate help to individuals in mental crises. The effectiveness of these hotlines largely depends on accurately identifying callers' emotional states, particularly underlying negative emotions indicative of increased suicide risk. However, the high demand for psychological interventions often results in a shortage of professional operators, highlighting the need for an effective speech emotion recognition model. This model would automatically detect and analyze callers' emotions, facil-itating integration into hotline services. Additionally, it would enable large-scale data analysis of psychological support hotline interactions to explore psychological phenomena and behaviors across populations. Our study utilizes data from the Beijing psychological support hotline, the largest suicide hotline in China. We analyzed speech data from 105 callers containing 20,630 segments and categorized them into 11 types of negative emotions. We developed a negative emotion recognition model and a fine-grained multi-label classification model using a large-scale pretrained model. Our experiments indicate that the negative emotion recognition model achieves a maximum F1-score of 76.96%. However, it shows limited efficacy in the fine-grained multi-label classification task, with the best model achieving only a 41.74% weighted F1-score. We conducted an error analysis for this task, discussed potential future improvements, and considered the clinical application possibilities of our study. All the codes are public available at: https://github.com/cz10914/psy_hotline_analysis.

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.

Acute intracerebral hemorrhage is a life-threatening condition that demands immediate medical intervention. Intraparenchymal hemorrhage (IPH) and intraventricular hemorrhage (IVH) are critical subtypes of this condition. Clinically, when such hemorrhages are suspected, immediate CT scanning is essential to assess the extent of the bleeding and to facilitate the formulation of a targeted treatment plan. While current research in deep learning has largely focused on qualitative analyses, such as identifying subtypes of cerebral hemorrhages, there remains a significant gap in quantitative analysis crucial for enhancing clinical treatments. Addressing this gap, our paper introduces a dataset comprising 222 CT annotations, sourced from the RSNA 2019 Brain CT Hemorrhage Challenge and meticulously annotated at the voxel level for precise IPH and IVH segmentation. This dataset was utilized to train and evaluate seven advanced medical image segmentation algorithms, with the goal of refining the accuracy of segmentation for these hemorrhages. Our findings demonstrate that this dataset not only furthers the development of sophisticated segmentation algorithms but also substantially aids scientific research and clinical practice by improving the diagnosis and management of these severe hemorrhages. Our dataset and codes are available at https://github.com/songchangwei/3DCT-SD-IVH-ICH.

The combined models of Graph Neural Network (GNN) and Recurrent Neural Network (RNN) are widely used for patient similarity computation. However, these studies mainly use the medical concepts to organize patient graphs, while a lot of concepts in Electronic Medical Records (EMRs) are paratactic, learning the temporal information based on concept sequences may introduce noise to similarity computation. To address this problem, we propose an Event Graph Learning Network (EGLN) to learn patient similarity. Specially, we firstly leverage the trained Event Extraction (EE) model to obtain the event elements. Then, aggregating the paratactic concepts of each medical event to construct the event graph for the patient to avoid the negative influence of nonexistent temporal information between paratactic concepts. Finally, the spatial and temporal semantic information between event nodes is aggregated for similarity computation. We evaluate EGLN leveraging a real-world dataset, and the experiment results indicate that our proposed EGLN model outperforms all baselines.

Accurate lesion segmentation on breast ultrasound (BUS) images is a crucial procedure in computer-aided ultrasonic diagnosis. Owing to the privacy of BUS data and the complexity of acquiring pixel-level labels, numerous researches attempt to achieve breast lesion segmentation with predefined feature-based and deep learning-based methods in a unsuper-vised or weakly supervised scenario. Although the former can typically extract more reliable contour information of the lesion, it is severely interfered by irrelevant tissues due to its inability to capture any semantic information. Furthermore, the weakly supervised deep learning segmentation based on class activation map (CAM) can explore the semantic information while failing to provide precise contour information. In view of the above observation, we present a weakly supervised framework merging complementary contour and semantic information for early lesion segmentation in BUS images. Specifically, guided by the prior knowledge of breast anatomy, we first extract and filter the contour information of suspected lesions located in breast parenchyma layer by clustering and morphological characteristic, respectively. Afterward, semantic information extraction is performed by a classification network to automatically explore the category information of the lesion. Finally, we selectively fuse the complementary information to facilitate lesion segmentation performance with more comprehensive features. Extensive experiments are conducted on the public dataset BUSI, and the results confirm the validity of our approach.

Biomedical named entity recognition (BNER) is a basic task of the extraction of medical information. The existing deep learning-based approaches usually represent the medical text by using words or characters. However, most of biomedical terms consist of many words (characters). Splitting them into many fragments (words or characters) while leveraging the attention mechanism to assign attention scores for each fragment maybe disperse the importance weight and cause a lower attention score for the biomedical terms. Therefore, this paper presents a dictionary-based adaptive attention network for BNER. Specifically, a biomedical dictionary is firstly constructed by integrating multiple existing medical resources. Secondly, building the guidance vectors by matching the electronic medical record (EMR) text to the constructed dictionary. Then, an adaptive attention strategy is presented to guide the attention mechanism to assign higher attention to the overall medical term by using the guidance vectors. We conduct extensive experiments on a real-world dataset, the results illuminate that our presented method outperforms all baselines.

Biomedical named entity recognition (BNER) is a critical task for biomedical information extraction. Most popular BNER approaches based on deep learning utilize words and characters as features to represent medical texts. However, many medical terminologies are composed of multiple words and characters, and splitting medical terminology into multiple words (or characters) and assigning weight values for each word (or character) by a standard attention mechanism may disperse the attention score and result in a lower weight value for the medical terminology. This paper proposes a Dictionary-guided Attention Network (DGAN) for BNER in Chinese electronic medical records (EMRs). First, the medical concepts are extracted as large-size words to supplement the comprehensive semantic information of the medical terminology by matching the EMR text to the biomedical dictionary. Then, based on the matched dictionary results, an optimized attention strategy is proposed to focus on the medical concept and adaptively assign higher weights to the characters contained in a concept. Furthermore, semisupervised learning is introduced to reduce the manual labeling of data and to handle the entities not defined in the medical dictionary. To validate our new model in recognizing biomedical named entities, we conduct comprehensive experiments on a real-world Chinese EMR dataset and the CCKS2017 dataset. Our promising results illustrate that our method not only achieves a state-of-the-art performance in BNER but also reduces manual data annotation.

The automatic disease diagnosis utilizing clinical data has been suffering from the issues of feature sparse and high probability of missing values. Since the graph neural network is a effective tool to model the structural information and infer the missing values, it is becoming the dominant method for the predictive model construction from electronic medical records. Existing graph neural network based solutions usually adopt the medical concepts (e.g., symptoms) the feature representation of clinical data without considering their underlying semantic relations. The limited discriminative capability of the medical concept cannot provide sufficient indicative information about the disease. This article proposes a knowledge-guided graph attention network for the disease prediction. Beside extracting the attribute-value structure as a large-size medical concept, the mutual information between multiple medical concepts mentioned in the electronic medical records are taken into account in the graph construction. Meanwhile, the defined diseases and their associations with the medical concepts in the medical knowledge graph are incorporated into the graph, which provides the potentials to enhance the indicative impacts of the medical concepts directly related to a target disease. Then, the spatial and attention based graph encoders are employed to aggregate information from directly neighbor nodes to generate node embeddings as the compact features to be used for disease diagnosis. The approach itself is a general one that can utilized to build the predictive model using Chinese EMRs for different diseases. The empirical experiments for its performance evaluation are conducted on the real-world COPD EMR dataset. The comparison study results show that the proposed model outperforms baseline methods, which illustrates the effectiveness of our proposed model.

Recently, lexicon-based Chinese Named Entity Recognition (NER) models have achieved state-of-the-art performance by benefiting from the rich boundary and semantic information contained in the lexicon. However, in the Chinese medical domain, it’s difficult to obtain the medical lexicon related to the target medical corpus. In this paper, we propose a new paradigm, enhancing Chinese medical NER with Auto-mined Lexicon (ALNER), which alleviates the difficulty of obtaining the medical lexicon by designing a data-driven automatic lexicon construction method. We define medical lexicon construction as a high-quality phrase mining task. We perform secondary annotation on the NER annotated data and use the secondary annotated data to train a deep learning-based phrase tagger. Experimental results show that our method can be combined with different lexicon-based Chinese NER models to improve performance and that the method does not require an external medical lexicon.

The widely deployed of hospital information systems causes an explosive growth of the electronic medical records (EMRs). It makes the medical structured processing technologies become critical to find researchable data in the large medical dataset. However, the high quality structured processing is a challenging task, in particular due to the inherent complexity and polysemy of medical terminology. In this paper, we propose a novel approach to achieve the joint extraction of events in Chinese electronic medical records, which solves the problem of cascading error transmission in traditional models and the ambiguity of Chinese characters. We first use the Bi-directional Encoder Representation from Transformers(BERT) model to mine features from the preprocessed medical data; then based on the characteristics of Chinese, we use the Bi-directional Long Short-Term Memory(BILSTM) model to capture the semantic information of the context. The experiments were conducted on a real dataset. The F1 score of our model in the identification and classification tasks of event triggers and arguments is the highest, reaching 71.6, 68.1, 55.4 and 46.9, respectively, which proves the effectiveness of the proposed method.

The integration of lexicon information into character-based models is a hot topic in Chinese Named Entity Recognition(NER) research. Most methods only utilize information from a single lexicon which is usually a general lexicon. However, In the Chinese medical text scenario, due to the large amount of medical terminology, a single lexicon, especially a general lexicon, offers little performance improvement to the Chinese NER. In this paper, we propose a Multi-source Lexicon Information Fusion method for Named Entity Recognition in Chinese Medical Text(MLNER) which can utilize information from both general and medical lexicons. Considering the small medical annotated corpus, we combine the model with the pre-trained model to improve the performance of the model on small datasets by exploiting the rich representation capability of the pre-trained model. Experiments show that our method can effectively improve the performance of NER in Chinese medical text. Our model is also applicable to Chinese NER tasks in other domain specific fields, with good scalability and application value.

The wide adoption of electronic medical record (EMR) systems causes rapid growth of medical and clinical data. It makes the medical named entity recognition (NER) technologies become critical to find useful patient information in the medical dataset. However, the medical terminologies usually have the characteristics of inherent complexity and ambiguity, it is difficult to capture context-dependency representations by supervision signal from a simple single layer structure model. In order to address this problem, this paper proposes a hybrid model based on stacked Bidirectional Long Short-Term Memory (BILSTM) for medical named entity recognition, which we call BSBC (BERT combined with stacked BILSTM and CRF). First, we use Bidirectional Encoder Representation from Transformers (BERT) to perform unsupervised learning on an unlabeled dataset to obtain character-level embeddings. Then, stacked BILSTM is utilized to obtain context-dependency representations through the multi hidden layers structure. Finally, Conditional Random Field (CRF) is used to predict sequence tags. The experiment results show that our method significantly outperforms the baseline methods, it serves as a strong alternative approach compared with traditional methods.