Foundational vision-language models (VLMs), such as CLIP, are emerging as a promising paradigm in vision tasks due to their strong generalization ability. Nevertheless, adapting them to downstream tasks remains challenging, especially in biomedical imaging, where scarce annotations, low-contrast features and complex patterns hinder model adaptation. Thus, prompt tuning is employed to facilitate the adaptation of VLMs. However, current prompt tuning methods like Context Optimization (CoOp) mainly learn a single yet static prompt which is applied to all images, and such one-size-fits-all prompt cannot describe the case-specific diagnostic cues in biomedical data, compromising the adaptation of VLMs. To this end, we propose a Dynamic Prompt Policy learning method that enables efficient adaptation of Biomedical VLMs (BioDPP) for accurate and highly generalizable few-shot biomedical image classification. Specifically, we conceptualize the learnable context as an agent, and present a paradigm of learning a dynamic prompting policy, rather than obtaining a single yet static prompt. Wherein, a dual-reward mechanism is developed to guide policy learning via the feedback on both classification decision and the consistency between the prompt and the context, steering the agent to generate context-aware prompts. Moreover, we devise adaptive baseline stabilization to dynamically regulate reward advantage value throughout the training process, enabling policy refinement in a complex reward space tailored to biomedical VLMs. Extensive experiments are conducted on 10 biomedical datasets, and the results reveal that our BioDPP achieves superior performance, demonstrating more efficient prompt optimization in biomedical VLMs.

Myocardial infarction remains a leading cause of cardiovascular mortality, making its early detection and localization via electrocardiograms (ECGs) critical for timely intervention. Despite advances in deep learning-based ECG analysis, two key challenges persist: (1) Severe class imbalance biases models toward dominant categories, hindering the learning of discriminative patterns for underrepresented subtypes (e.g., posterior, extensive anterior infarctions). (2) The inability to adequately model complex spatiotemporal dependencies in ECGs limits the effectiveness of current approaches. To address these challenges, we propose SPG-Net, a novel structure-aware spatiotemporal graph network that integrates clinical priors with dynamic spatiotemporal dependencies of ECG for MI localization. The framework extracts high-resolution lead-wise features and constructs anatomically grounded graphs using class-specific lead contribution maps. A dedicated graph attention module integrates static anatomical priors, inter-lead spatial interactions, and temporal dependencies to generate ECG feature representations that preserve physiological consistency while accommodating individual variability. Furthermore, we design a decoupled classification mechanism to reduce interference from overlapping labels, significantly improving minority-class recognition. Comprehensive experiments on PTB and PTB-XL datasets demonstrate SPG-Net's superior overall performance and significantly enhanced minority-class recognition, validating its efficacy in class-imbalanced MI localization.

Automatic summarization of discharge summaries aims to generate a concise yet informative overview from complex electronic health records (EHRs), thus alleviating the documentation burden on clinicians. Existing methods usually employ pre-trained Transformer architectures to capture textual dependency via self-attention mechanism. However, these methods fail to utilize the information of multi-level hierarchical structure in EHR, compressing the understanding of intricate semantic interaction among different structure levels. In this work, we propose HiGraphSum, a hierarchical hypergraphbased dual-stage model for automatic summarization of discharge summaries. In the first stage, HiGraphSum constructs a hierarchical hypergraph that integrates heterogeneous wordsentence graphs and sentence-based hypergraphs to capture hierarchical and contextual information, facilitating global document structure and intricate clinical semantics learning. In the second stage, we introduce a probability concentration strategy (PCS) to refine summary generation by sharpening the output distribution, thereby enhancing the factual consistency of the generated summaries. Extensive experiments on discharge summarise datasets demonstrate that HiGraphSum outperforms state-of-the-art baselines.

Automatic text summarization in the biomedical field can support efficient literature screening, medical knowledge management, and innovative medical research. In recent years, Large Language Models (LLMs), as a disruptive technology in natural language processing, have shown great potential for Biomedical Automatic Text Summarization (BATS). This technology helps to better understand the terminology of biomedical texts, track medical hotspots, and generate personalized diagnoses and treatment plans. This paper provides an in-depth discussion on the development of BATS, and the opportunities as well as challenges brought by applying LLMs to biomedical automatic text summarization. Firstly, the development of BATS is reviewed, where traditional text summarization, neural network-based summarization, and LLMs-based summarization are analyzed systematically. Meanwhile, the applications of various LLMs (e.g., BERT and GPT series) in three types of BATS are presented in detail, including extractive summarization, abstractive summarization, and hybrid summarization. Next, the relevant datasets are introduced, such as PubMed, COVID-19 and MIMIC-Ⅲ. Then, traditional, emerging, and auxiliary metrics for evaluating the performance of BATS are shown, and the performance evaluation of different models is elaborated. Finally, the opportunities brought by applying LLMs to BATS are described, and the potential challenges along with the corresponding solutions are discussed.

Effective diabetes management requires regular and accurate blood glucose monitoring; however, traditional invasive methods often cause discomfort and inconvenience. Non-invasive techniques such as photoplethysmography (PPG) have been explored, though single-wavelength PPG systems are limited by the overlapping absorption characteristics between glucose and other biological components, such as water and fat. In this study, a novel multi-wavelength PPG system integrated with temperature and humidity sensors is introduced, coupled with a neural network framework featuring attention mechanisms to enhance glucose prediction. The system employs six optical sensors covering wavelengths from the visible to near-infrared (NIR) spectrum, enabling deeper tissue penetration and enhanced glucose specificity by targeting distinct absorption peaks-especially those above 1000 nm. The system was validated using a robust dataset of 26,063 measurements from 254 participants. The experimental results demonstrate significant improvements, with the model achieving 86.49% compliance with the ISO 15197: 2013 standards and 91.80% of measurements falling within Zone A of the Parkes error grid. The introduction of multiple wavelengths clearly improves performance over single-wavelength systems, and wavelengths above 1000 nm were shown to have a higher contribution in glucose prediction. In addition, the incorporation of temperature and humidity data also enhanced performance by accounting for environmental and physiological factors, and that demographic and meal-related factors significantly impact prediction accuracy, thereby underscoring the potential of this system as a reliable, non-invasive, and personalized glucose monitoring tool.

It is crucial to predict the in-hospital mortality for improving clinical decision-making and optimizing hospital resource allocation. Many recent studies have attempted to integrate multimodal data, such as digital time series and clinical notes from electronic health records (EHRs), to improve the performance of mortality prediction. Although current methods show good performance, it is often difficult to effectively capture fine-grained inter-modal correspondences. In addition, the different distributions and heterogeneous properties of the various modalities lead to modality gaps that severely affect the effectiveness of modal fusion. To overcome these limitations, we propose an Attention-based Multimodal Fusion with Adversarial network (AMFA) for in-hospital mortality prediction. In AMFA, feature extraction is first performed to obtain modality-specific features, and then an attention mechanism is employed to capture inter-modal interrelationships. The main difference with existing methods is that AMFA achieves this by calculating the relative importance of individual features focusing attention on the most relevant features and eliminating irrelevant features, and then reassigning attention between relevant features to obtain finer semantic relevance. Subsequently, we introduce a discriminator network that addresses the modality gap by adjusting the distribution of various modal representations through adversarial training. We evaluate AMFA on two publicly available datasets, and experimental results show that AMFA outperforms several state-of-the-art models in the task of in-hospital mortality prediction.

Predicting future health status based on historical patient visits is one of the essential tasks in healthcare. Many existing approaches attempt to enhance the representation learning capability of models by incorporating relevant medical knowledge, but their effectiveness is severely affected by the incompleteness and noise of the knowledge graphs. Moreover, due to the inability to capture temporal features at a fine-grained level, most existing methods also have limitations in learning the temporal development of patients’ health status. To address these issues, we propose a Knowledge-Enhanced Difference-Aware clinical time series representation learning model (KEDA) for diagnosis prediction. In this model, we first combine the medical ontology graph and co-occurrence graph, and use hierarchical graph convolution and contrastive learning methods to enhance the semantic representation of medical entities. After that, a task-specific difference-aware temporal module is designed to improve the accuracy of patient representation, which adds two novel gated units in the original GRU to fuse multi-type clinical information based on the relationship between different types of data and prediction tasks and capture fine-grained temporal evolution of patient health status. We validate our model on two publicly available datasets, and the experimental results demonstrate that KEDA outperforms the state-of-the-art methods.

Triplet extraction is one of the fundamental tasks in biomedical text mining. Compared with traditional pipeline approaches, joint methods can alleviate the error propagation problem from entity recognition to relation classification. However, existing methods face challenges in detecting overlapping entities and overlapping relations, which are ubiquitous in biomedical texts. In this work, we propose a novel pipeline method of end-to-end biomedical triplet extraction. In particular, a span-based detection strategy is used to detect the overlapping triplets by enumerating possible candidate spans and entity pairs. The strategy is further used to capture different contextualized representations via an entity model and a relation model, respectively. Furthermore, to enhance interrelation between spans, entity information from the output of the entity model is used to construct the input for the relation model without utilizing any external knowledge. Our approach is evaluated on the drug-drug interaction (DDI) and chemical-protein interaction (CHEMPROT) datasets, exhibiting improvement of the absolute F1-score in relation extraction by 3.5%-3.7% compared prior work. The experimental results highlight the importance of overlapping triplet detection using the span-based approach, acquisition of various contextualized representations via different in-domain pre-trained language models, and early fusion of entity information in the relation model.

Disease prediction based on the Electronic Health Record (EHR) is an important task in healthcare. EHR records patients’ every visit by time, and there are many kinds of medical codes within a visit, therefore, EHR has characteristics of temporal irregularity and hierarchical structure. Some recent works employ BERT-style models to process EHR data for disease prediction. However, few of these models can give consideration to capture both the interaction between medical codes and the impact of temporal irregularity. To solve this problem, we propose the Double-Level BERT-style model (DL-BERT). Considering EHR’s hierarchical structure, the model contains a code-level and a visit-level representation layer which can learn the relationship between medical codes and temporal influence respectively. In the code-level representation layer, the model achieves the representation power by employing external medical ontologies to provide multi-resolution information of medical codes and the Transformer to embed medical codes. Besides, the model adopts two pre-training tasks to enhance the ability to capture the link between different kinds of codes. In the visit-level representation layer, DL-BERT utilizes a special time-aware Transformer to model temporal information. And the model adopts a visit-level pre-training task for better learning context information. Experiments are conducted on two real-world healthcare datasets and show that our model outperforms all baselines demonstrating the effectiveness of our model.

Early detection of paroxysmal atrial fibrillation (AF) is valuable in determining treatment options and diagnosing complications, and identifying the number and duration of AF episodes in the dynamic electrocardiogram (ECG) may help advance the study of pathological AF mechanisms. Current computer-aided ECG signal diagnosis methods can be divided into two categories: (a) methods based on every single heartbeat split from ECG records, which cannot use information from adjacent heartbeats, and (b) methods based on a segment of ECG records that contain multiple heartbeats, which can only give a classification of the whole segment but cannot identify the categories of each heartbeat in the record at a fine granularity. Few methods can exploit the relationship between heartbeats as well as identify the onset and end of disease episodes. To fill this gap, we designed an improved U-Net model named Percept U-Net to detect AF episodes from the ECG records. In this model, we propose the percept attention to replace skip connections in U-Net to integrate low-level features with high-level features, thereby enhancing the ability to detect and discriminate between normal and abnormal heartbeats. Experimental results show that Percept U-Net achieves higher classification and localization accuracy with less computational overhead and paraments than other comparative methods.

Nuclei segmentation is valuable in histopathological image analysis, but labeling nuclei is costly. Different organs, pa-tients and diseases will lead to high variability in the morphology of nuclei, the structure of tissues, etc., which is difficult to elim-inate. Inconsistent staining operations and scanning operations will cause variability in histopathological image style. Relying on a small amount of labeled data, it is hard for the model to adapt to the high variability among histopathological images. Therefore, it is necessary to exploit the value in the massive unlabeled data. However, because the existing pretext tasks in self-supervised learning do not well consider the characteristics of histopathological images and segmentation task, the same for the existing data augmentation approaches in contrastive learning, they are not suitable for nuclei segmentation. In this paper, the proposed method, named NormToRaw, takes into consideration the characteristics of nuclei segmentation, which can learn semantic information from different stains by style transfer. A generative adversarial network is used to transfer the normalized image to the raw image. Pre-trained on more than 8,000 unlabeled images and trained on 16 labeled images, the experimental results of 5 pre-trained models showed that the proposed method is effective for improving the performance of nuclei segmentation.

Predicting the future diagnoses from patients’ historical Electronic Health Records (EHR) is a significant task in healthcare. EHR consist of multiple modal data, each modality has different features and contains a wealth of information of patients. However, most of the existing EHR-based prediction methods either only use unimodal data, or fail to fully explore the correlation between different modalities when fusing multimodal data. To address these challenges, we propose a Multimodal Attention-based fusIon Networks (MAIN) for diagnosis prediction. In this model, we first design different feature extraction modules for each modality. Then, an inter-modal correlation module which contains two layers is applied to capture the intermodal correlation. Finally, a multimodal fusion module based on weighted averaging is utilized to integrate the representations derived from different modalities and their correlation to obtain the patient representation for diagnosis prediction. We evaluate our proposed model on two medical datasets, and the experimental results demonstrate the effectiveness of MAIN.

Compared with the general complex network, the multilayer network is more suitable for the description of reality. It can be used as a tool of network pharmacology to analyze the mechanism of drug action from an overall perspective. Combined with random walk algorithm, it measures the importance of nodes from the entire network rather than a single layer. Here a four-layer network was constructed based on the data about the action process of prescriptions, consisting of ingredients, target proteins, metabolic pathways and diseases. The random walk algorithm was used to calculate the betweenness centrality of the protein layer nodes to get the rank of their importance. According to above method, we screened out the top 10% proteins that play a key role in treatment. Prescriptions Xiaochaihu Decoction was taken as example to prove our method. The selected proteins were measured with the ones that have been validated to be associated with the treated diseases. The results showed that its accuracy was no less than the topology-based method of single-layer network. The applicability of our method was proved by another prescription Yupingfeng Decoction. Our study demonstrated that multilayer network combined with random walk algorithm was an effective method for pre-screening vital target proteins related to prescriptions.

High-risk prediction of cardiovascular disease is of great significance and impendency in medical fields with the increasing phenomenon of sub-health these years. Most existing pathological methods for the prognosis prediction are either costly or prone to misjudgement. Therefore, plenty of automated models based on machine learning have been proposed to predict the onset of cardiovascular disease with the premorbid information of patients extracted from their historical Electronic Health Records (EHRs). However, it is a tough job to select proper features from longitudinal and heterogeneous EHRs, and also a great challenge to obtain accurate and robust representations for patients. In this paper, we propose an entirely end-to-end model called DeepRisk based on attention mechanism and deep neural networks, which can not only learn high-quality features automatically from EHRs, but also efficiently integrate heterogeneous and time-ordered medical data, and finally predict patients' risk of cardiovascular diseases. Experiments are carried out on a real medical dataset and results show that DeepRisk can significantly improve the high-risk prediction accuracy for cardiovascular disease compared with state-of-the-art approaches.