Health Information Processing

With the rise of social media, users increasingly share their medication experiences online, offering a new data source for real-time ADR detection. However, ADR detection on social media faces two key challenges: limited labeled data and an imbalance between positive and negative samples. While previous studies have explored solutions like transfer learning, multi-source data fusion, joint task training, and loss function optimization, these approaches can introduce noise, increase annotation costs, or complicate training complexity. Moreover, despite the promising zero-shot and few-shot capabilities of large language models (LLMs) in natural language processing tasks, their performance in social media-based ADR detection remains below that of smaller, fine-tuned models. To tackle these challenges, we propose the Bal-LLaMA framework, comprising three modules: a data augmentation module to balance positive and negative samples as well as mitigating the challenge posed by limited annotated data, an instruction data construction module tailored for social media ADR detection, and a QLoRA-based module for efficient parameter fine-tuning. Experimental results demonstrate that Bal-LLaMA significantly outperforms existing state-of-the-art models on various social media ADR detection datasets, confirming the effectiveness of our approach.

Assessing discharge conditions for elderly coronary patients, especially those with serious or complex disease symptoms, is a crucial task for physicians. Clinical evaluations need to consider demographic information, physiological monitoring, diagnoses, and vital signs, which are multi-dimensional, multi-source, and heterogeneous data. While machine learning models can predict discharge risk based on these data, challenges remain in dealing with redundant features and extracting key features to enhance prediction performance.

In this work, we restructured patients’ comprehensive characteristics into a data sequence according to multiple vital clinical stages. Through feature engineering, we constructed steady-state indexes (SSIs), which are generated features that track the patient’s condition changes and the stability of vital signs. Additionally, we standardized variable-length biochemical test data to a fixed-length to address the issue of inconsistent data lengths. Then, we proposed a two-stage multi-module deep fusion network for discharge risk assessment. In the first stage, the data was divided into modules and we extracted features from biochemical test data, SSIs, and comprehensive clinical data using a transformer encoder, CNN, and BiLSTM, respectively. In the second stage, we designed a dual-layer attention fusion network, where dual pooling channel attention was applied to biochemical test data to capture more relevant relationships for discharge results, and sparse attention combining local and simple global information was used on aggregated features to reduce computational complexity.

Experiments were conducted on datasets collected from three local 3A hospitals, and the results demonstrated that our method outperformed other methods in the evaluated metrics. Ablation experiments further verified the benefits of segmenting different types or sources of data into different modules for clinical data analysis.

Objective: To analyze the clinical data of children with acute liver failure to identify risk factors for hemorrhagic complications, thereby providing a basis for clinical diagnosis and treatment decisions in pediatric acute liver failure (PALF). Methods: Clinical data from children diagnosed with acute liver failure and hospitalized at the Children’s Hospital affiliated with Chongqing Medical University from January 2014 to June 2024 were collected. Data included general information, laboratory indicators, and hemorrhagic complications. Patients were categorized into hemorrhagic and non-hemorrhagic complication groups for comparative analysis. Results: A total of 663 cases were analyzed, with 239 cases (36.05%) having hemorrhagic complications and 424 cases (63.95%) without. Only 21 cases (3.17%) had spontaneous bleeding. Multivariate analysis identified infection (OR = 4.05, 95%CI: 2.47~6.64, p < 0.001), hepatorenal syndrome (HRS) (OR = 2.95, 95%CI: 1.82~4.77, p < 0.001), multiple organ dysfunction syndrome (MODS) (OR = 2.47, 95%CI: 1.57~3.88, p < 0.001), low platelet count (≤50 × 10^9/L) (OR = 2.51, 95%CI: 1.58~3.99, p < 0.001), and low fibrinogen (≤1 mg/dL) (OR = 1.73, 95%CI: 1.12~2.69, p = 0.014) as independent risk factors for hemorrhagic complications in children with acute liver failure. Conclusion: Spontaneous bleeding in PALF is relatively rare, with infection, HRS, MODS, thrombocytopenia, and hypofibrinogenemia being high-risk factors for bleeding. Active prevention and control of infections and maintenance of organ function are crucial in preventing bleeding in PALF. Routine coagulation tests like prothrombin time and international normalized ratio are not sufficient to predict bleeding risk in PALF, necessitating a more comprehensive coagulation evaluation system.

Obstructive sleep apnea (OSA) is a common sleep-disordered breathing (SDB) characterized by recurrent apnea events during sleep due to partial or complete obstruction of the upper airway, which impairs the patient’s quality of sleep and daily life and increases the risk of several chronic diseases. Polysomnography (PSG) is clinically used as the gold standard for detecting OSA, but its expensive, complex and time-consuming procedure limits its widespread use. In this study, we propose a novel pseudo-modal fusion network (PMFNet) using single-lead ECG signals for the task of obstructive sleep apnea. We innovatively propose the concept of “pseudo-modal” in ECG signal analysis, using the QRS wave detection algorithm and the continuous wavelet transform (CWT) to obtain the R-wave data features and time-frequency maps as pseudo-modal data, respectively. Different feature extractors are carefully designed for these two types of pseudo-modal data to complete the time-domain and frequency-domain feature extraction of ECG signals, and a bilinear attention network (BAN) is introduced to effectively integrate our proposed pseudo-modal data. We validate the performance of PMFNet on the publicly available PhysioNet Apnea-ECG dataset and make a fair comparison with previous studies. Extensive experimental results show that PMFNet achieves optimal performance, with specific accuracy, sensitivity, specificity, and F1 scores of 92.79%, 90.90%, 93.95%, and 90.42%, respectively. Our proposed PMFNet is able to determine end-to-end whether an OSA event has occurred for a given ECG signal segment, providing an effective and convenient alternative to current clinical approaches.

The early detection of glottic carcinoma is critical for improving patient outcomes, as it enables timely intervention, preserves vocal function, and significantly reduces the risk of tumor progression and metastasis. However, the similarity in morphology between glottic carcinoma and vocal cord dysplasia results in suboptimal detection accuracy. To address this issue, we propose a vision large language model-based (VisionLLM-based) multimodal fusion network for glottic carcinoma detection, known as MMGC-Net. By integrating image and text modalities, multimodal models can capture complementary information, leading to more accurate and robust predictions. In this paper, we collect a private real glottic carcinoma dataset named SYSU1H from the First Affiliated Hospital of Sun Yat-sen University, with 5,799 image-text pairs. We leverage an image encoder and additional Q-Former to extract vision embeddings and the Large Language Model Meta AI (Llama3) to obtain text embeddings. These modalities are then integrated through a laryngeal feature fusion block, enabling a comprehensive integration of image and text features, thereby improving the glottic carcinoma identification performance. Extensive experiments on the SYSU1H dataset demonstrate that MMGC-Net can achieve state-of-the-art performance, which is superior to previous multimodal models.

The rapid advancements of large language models (LLMs) have opened new avenues for data processing and knowledge extraction, particularly in the medical domain. This paper investigates the application of LLMs in Traditional Chinese Medicine (TCM), with a focus on enhancing the models’ capabilities in syndrome differentiation thinking tasks. We propose a method that delineates the syndrome differentiation process in TCM into four critical steps: clinical information extraction, pathogenesis inference, syndrome inference, and explanatory summarization, with tailored prompting strategies designed for each step. By integrating Retrieval-Augmented Generation (RAG) with instruction tuning, we generated 800 instruction data entries rich in localized knowledge and instruction tuning of a pre-trained model. Experimental results indicate that our approach significantly improves the models’ performance in TCM syndrome differentiation thinking, achieving top rankings in both the A and B leaderboards, with scores of 45.12 and 44.37, respectively.

Drug-target interaction prediction help reduce the cost and time of drug development. However, existing research often overlooks the complexity of biological interactions. To address this issue, this paper proposes a DTI prediction model that integrates multi-dimensional biochemical features. Specifically, the model utilizes a heterogeneous graph attention network to capture the topological relationships among biological entities from diverse data, providing a deep understanding of the interactions among drugs, proteins, diseases, and side effects. A molecular attention Transformer network and a CBiNet module are used to extract the key structural features of drugs and targets. By automatically optimizing the weight distribution between drugs and targets, the model enhances the information transfer during network training, significantly improving model performance. Experimental results on real-world datasets demonstrate that the proposed model outperforms the current state-of-the-art methods in the field. Among the top 50 novel COVID-19 therapeutic drugs predicted by our model, 30 have been supported by clinical trials or scientific literature, further demonstrating the effectiveness of our proposed method in drug repurposing. Experimental data and supplementary materials are available online at: https://​github.​com/​Eadog/​MBF-DTI.

Structure-based drug design (SBDD) accelerates drug discovery but traditionally relies on labor-intensive, simulation-based methods. Deep generative models offer a data-driven alternative, but the currently prevalent ligand-based models are often constrained by the availability of active compounds. Here, we present a new SO(3)-equivariant generative model for SBDD, using a pseudo-ligand point-cloud representations of protein cavities to optimize ligands and generate stable 3D molecules. Our model accurately models the chemical space of the protein-binding compounds. We evaluated our model on three therapeutic targets: Janus kinase 2 (JAK2), peptidylprolyl isomerase (hPin1), and Mycobacterium tuberculosis malate synthase. Our model successfully rediscovered moieties involved in key interaction with the proteins and proposed alternative moieties with bioactivity supported by the literature among the highly ranked generated samples. This work approach offers a new way to ligand optimization and drug discovery, advancing the field of public health science by enhancing the precision and efficiency of molecular design in therapeutic development.

Recent studies indicate that variations in the expression of circular RNA (circRNA) can alter cellular drug sensitivity, which in turn significantly impacts drug efficacy and plays a crucial role in human health and disease treatment. Thus, predicting the associations between circRNA and drug sensitivity is essential. In this study, to enhance the prediction accuracy, we introduce a novel method named MHCDA that leverages graph convolutional networks and hypergraph convolutional networks to extract both local and global information of the circRNA-drug network. Specifically, MHCDA first constructs homogeneous graphs for circRNAs and drugs through similarity fusion networks, then obtain the representations of circRNAs and drugs using graph convolutional networks. On the other hand, we utilize hypergraph convolutional networks to extract more complex higher-order interactions between drugs and circRNAs, respectively. Utilizing Contrastive Learning to Analyze circRNA Feature Representations and Pharmacological Feature Representations Across Various Convolutional Architectures. Meanwhile, we utilize autoencoders to extract circRNA and drug features from the established associations between circRNAs and drugs. Finally, we integrate the various features obtained to predict the relationship between circRNA and drug sensitivity. Experiments reveal that the AUC and AUPR values of MHCDA are 0.918 and 0.929, respectively, surpassing those of other advanced models.

The purpose of relational extraction is to extract the relational triples in a given text. Recent work shows that Large Language Models(LLMs) achieve excellent performance in information extraction, especially in the few-shot learning. An important challenge in the medical field is the long-tail problem, which has not received much attention in the context of LLMs. Therefore, we propose a novel approach, the Knowledge Enhancement-Prompted Relationship Extraction (KE-PromptRE), to improve the performance of LLM for long-tail problems of relationship extraction in the medical domain. KE-PromptRE transforms labeled data in the medical domain into generative data accepted by LLM and infuses it into LLM to enhance LLM’s performance in the medical domain. In addition, we use Curriculum Learning (CL) to infuse knowledge data into LLM in batches according to difficulty to improve the performance of LLM on the relation extraction. The results showed that our approach performed 9.5% and 2.3% above baseline on the Critical Illness entities and relationships Corpus (CIC) and Chinese Medical Information Extraction (CMeIE) datasets.

Drug-target interaction prediction can help researchers understand the mechanism of action of drugs and discover new drug targets, and even assist researchers in designing more effective drug therapeutic regimens, which will be of great significance to drug development. In recent years, computer-aided technology has been better applied in various fields, and the paper will discuss drug-target interaction prediction methods: molecular docking, ligand-based, text mining, and feature-based methods in the context of computer-aided technology. The paper mainly discusses and elaborates on the feature-based techniques and analyzes each type of method’s principles, advantages, and disadvantages. At the same time, the paper also discusses the problems and challenges faced by drug-target interaction prediction, including its dataset, cold-start, and model design problems. Finally, deep learning technology will still play a great potential for application in this field and briefly point out the direction and trend of future research.

Joint entity-relation extraction is a critical task in transforming unstructured or semi-structured text into triplets, facilitating the construction of large-scale knowledge graphs, and supporting various downstream applications. Despite its importance, research on Chinese text, particularly with complex semantics in specialized domains like medicine, remains limited. To address this gap, we introduce the CH-DDI, a Chinese drug-drug interactions dataset designed to capture the intricacies of medical text. Leveraging the strengths of attention mechanisms in capturing long-range dependencies, we propose the SEA module, which enhances the extraction of complex contextual semantic information, thereby improving entity recognition and relation extraction. Additionally, to address the inefficiencies of existing methods in facilitating information exchange between entity recognition and relation extraction, we present an interactive fusion representation module. This module employs Cross Attention for bidirectional information exchange between the tasks and further refines feature extraction through BiLSTM. Experimental results on both our CH-DDI dataset and public CoNLL04 dataset demonstrate that our model exhibits strong generalization capabilities. On the CH-DDI dataset, our model achieves an F1-score of 96.73% for entity recognition and 78.43% for relation extraction. On the CoNLL04 dataset, it attains an entity recognition precision of 89.54% and a relation extraction accuracy of 71.64%.

In the field of Knowledge graph Question Answering (KGQA), Semantic Parsing-based (SP) methods have become increasingly prominent. These methods, particularly those translating natural language into logical forms via generative models, have shown promising results. However, a key challenge in SP-based KGQA is the potential for noise introduction when incorrect or irrelevant information is used during the learning process. This noise can significantly degrade the performance of logical form generation, a critical aspect of KGQA. To tackle this issue, we propose a framework named the Multi-task with Loss Optimization for KGQA (MLO-KGQA), which significantly employs the balanced uncertainty weight loss approach to optimize the loss function in multi-task learning. MLO-KGQA takes logical form generation as the primary task, with entity disambiguation and subgraph selection as subtasks. The critical innovation of our framework is the application of balanced uncertainty weighting, which optimizes loss weights during multi-task learning, effectively reducing the noise problem. Experimental results on Pediatric Epilepsy Knowledge Graph Question Answer (PEKGQA) and CCKS2023-CKBQA show that the MLO-KGQA demonstrates a significant improvement in performance.

Electronic health records (EHR) store rich data of medical entities, such as diagnoses, procedures, and medications, which are invaluable in the development of automated systems for hypertension medication recommendations. These entities within EHR demonstrate significant synergies during the treatment process. However, existing medication recommendation methods predominantly focus on homogeneous graphs, thus overlooking the crucial synergistic relationships among heterogeneous medical entities. Moreover, accurately modeling the progression of hypertension using EHR is essential for precise medication recommendations, but current approaches often lack comprehensive temporal modeling and do not fully meet clinical requirements. To overcome these challenges, this paper introduces a novel model for hypertension medication recommendation that leverages the synergy and selectivity of heterogeneous medical entities. Initially, patient EHRs are utilized to construct both heterogeneous and homogeneous graphs. The inter-entity synergies are then captured using a multi-head graph attention mechanism, which enhances the entity-level representations. Subsequently, a dual-layer temporal selection mechanism calculates selective coefficients between current and historical visit records, thereby aggregating these to form refined visit-level representations. Ultimately, medication recommendation probabilities are determined based on these comprehensive patient representations, yielding practical and actionable recommendations. Experimental evaluations conducted on the real-world dataset MIMIC-IV v2.2 demonstrate that our model significantly outperforms baseline models. It achieves a Jaccard similarity coefficient of 55.82%, a precision-recall AUC of 80.69%, and an F1 score of 64.83%, thereby demonstrating its superior efficacy in medication recommendation. These results underscore the potential of our model to enhance clinical decision-making in the management of hypertension.

This paper addresses the challenge of integrating Traditional Chinese Medicine (TCM) principles with contemporary artificial intelligence to generate accurate and personalized dietary recommendations. Focusing on the TCM concept of “One Root of Medicine and Food,” we develop a novel method that employs Retrieval-Augmented Generation (RAG) techniques based on Large Language Models (LLMs). We confront the difficulties of translating nuanced TCM wisdom into actionable advice compatible with AI systems, ensuring high accuracy and relevance in personalized recommendations, and maintaining scientific rigor while preserving traditional knowledge. To overcome these obstacles, we design a unified set of prompt engineering instructions tailored for TCM dietary guidance and evaluate several mainstream LLMs, ultimately selecting Qwen as the optimal base model. By integrating RAG with a specialized TCM knowledge base, we enhance the model’s accuracy and professionalism; experimental results show significant improvements, with the ROUGE-L score increasing from 0.294 to 0.427 and the Accuracy score rising from 0.315 to 0.439. Case studies further demonstrate that our method enhances the rationality and customization of recommendations, ensuring they are scientifically sound and tailored to individual needs. This approach significantly improves the relevance and fidelity of TCM-based dietary recommendations, bridging traditional wisdom and modern technology for personalized healthcare.

The precision of contextual information is crucial for the results of Large Language Models (LLMs). However, to achieve more precise results in the field of medication instructions, which are characterized by their specificity in covering multiple drug characteristics, it is necessary to refine these instructions. Additionally, most current methods for constructing indexes do not consider sentence-level semantic information, which can easily ignore important details. Therefore, we propose a retrieval enhancement method based on ontology subdivision for LLM(OAGLLM). In the domain of Medication Instructions, we propose an ontology-based subdivision approach for constructing a specialized ontology and systematically storing expert knowledge, accurately match various aspects of Medication Instructions. To capture sentence-level semantic information and improve retrieval accuracy, we have introduced a hierarchical construction indexing method, which is designed to enhance text relevance and coherence. Lastly, we develop a retrieval augmentation system that integrates the Medication Instructions ontology with the hierarchical database. To validate the effectiveness of OAGLLM, we constructed a dataset on medication instructions. Our experiments demonstrate that our method outperforms other models in overall performance and excels across various types of data.