Herbal prescriptions are a cornerstone of Traditional Chinese Medicine (TCM), bridging diagnostic reasoning and therapeutic interventions. Existing herbal prescription recommendation (HPR) methods often rely on structured inputs or multi-stage pipelines, limiting their applicability to real-world clinical scenarios. To address this, we propose MKE-DAN (Medical Knowledge Enhanced Dual-level Alignment Network), a novel framework that directly generates herbal prescriptions from raw, unstructured medical records. By integrating a domain-specific pre-trained language model and label-wise self-attention, our model leverages intra-modal alignment to capture subtle diagnostic patterns and treatment reasoning from medical records. Additionally, cross-modal alignment with a heterogeneous graph transformer and knowledge-enhanced cross-attention ensures effective integration of structured knowledge with medical records for accurate and interpretable herb recommendations. Using TCM-9K, a large-scale dataset of clinical records paired with expert-verified prescriptions, and the CCL2025-800 dataset for external validation, experimental results demonstrate that MKEDAN achieves state-of-the-art performance, robust generalization, and strong handling of rare prescription patterns, showcasing its potential for real-world deployment in intelligent TCM systems.

Automated International Classification of Diseases (ICD) coding assigns standardized codes to clinical records, playing a critical role in healthcare systems. However, existing methods struggle with semantic gaps between clinical text and ICD codes, poor performance on rare codes, and limited interpretability. We propose TraceCoder, a framework that integrates multi-source external knowledge, including UMLS, Wikipedia, and large language models (LLMs), to enrich ICD code representations and provide traceable, evidence-based predictions. A hybrid attention mechanism is introduced to model interactions among labels, clinical context, and knowledge, improving longtail code recognition and interpretability by grounding predictions in external evidence. Experiments on MIMIC-III-ICD9, MIMIC-IV-ICD9, and MIMIC-IV-ICD10 datasets demonstrate that TraceCoder achieves state-of-the-art performance, with ablation studies validating its components. TraceCoder offers a scalable, interpretable, and reliable solution for automated ICD coding, aligning with clinical needs for accuracy and traceability.

Electronic medical records (EMRs) are crucial for modern healthcare, containing rich information about patient care, diagnoses, and treatments. However, their unstructured nature, domain-specific language, and complexity pose significant challenges for automated understanding. Existing methods often treat all data equally, limiting their ability to handle rare or complex cases effectively. We present TACL (Threshold-Adaptive Curriculum Learning), a novel framework that dynamically adjusts the training process based on sample complexity. Inspired by progressive learning, TACL categorizes data into difficulty levels, focusing on simpler cases early in training and gradually addressing more complex ones. A domain-specific pre-trained language model is used for difficulty assessment, considering semantic, syntactic, and contextual features. Additionally, TACL employs an adaptive training strategy to enhance task-specific performance and ensure generalization across diverse datasets. Experimental results on multilingual datasets, including MIMICIII, MIMIC-IV, and Chinese clinical records, demonstrate TACL's effectiveness in tasks such as ICD coding, readmission prediction, and TCM syndrome differentiation. TACL improves performance on rare and complex cases, providing a scalable and robust solution for medical text understanding.

Intraoperative hypotension (IOH), a critical condition characterized by sustained arterial pressure below 65 mmHg, poses significant risks of organ damage and mortality. Despite advancements in predictive models, current methods often fall short in capturing the dynamic, multi-scale nature of physiological signals due to their reliance on fixed convolutional kernels. To address this challenge, we introduce the Adaptive Time-Shift Convolutional Neural Network (ATS-CNN), a novel approach that combines the flexibility of deformable convolutions with the temporal modeling power of LSTMs. Unlike traditional methods, ATS-CNN leverages an innovative LSTM-based offset generation mechanism to dynamically adjust convolutional sampling points, enabling precise adaptation to the non-stationary characteristics of biosignals. Through a combination of adaptive sampling, deformable 1D convolutions, and robust data augmentation techniques-including dynamic noise injection-the model achieves unparalleled robustness and predictive accuracy. Extensive internal and external validations on two large-scale real-world datasets demonstrate that ATSCNN not only sets a new benchmark in$\mathbf{I O H}$prediction but also offers a clinically interpretable framework, transforming how machine learning addresses high-stakes medical challenges. This fusion of adaptability, precision, and interpretability positions ATS-CNN as a disruptive innovation in the field of perioperative care.

Intraoperative hypotension (IOH) is strongly associated with postoperative complications, including postoperative delirium and increased mortality, making its early prediction crucial in perioperative care. While several artificial intelligence-based models have been developed to provide IOH warnings, existing methods face limitations in incorporating both time and frequency domain information, capturing short- and long-term dependencies, and handling noise sensitivity in biosignal data. To address these challenges, we propose a novel Self-Adaptive Frequency Domain Network (SAFDNet). Specifically, SAFDNet integrates an adaptive spectral block, which leverages Fourier analysis to extract frequency-domain features and employs self-adaptive thresholding to mitigate noise. Additionally, an interactive attention block is introduced to capture both long-term and short-term dependencies in the data. Extensive internal and external validations on two large-scale real-world datasets demonstrate that SAFDNet achieves up to 97.3% AUROC in IOH early warning, outperforming state-of-the-art models. Furthermore, SAFDNet exhibits robust predictive performance and low sensitivity to noise, making it well-suited for practical clinical applications.

Timely prediction and intervention for Intraoperative Hypotension (IOH), a prevalent complication associated with general anesthesia, is crucial to prevent severe postoperative outcomes. While existing machine learning methods for IOH prediction have shown promise, they face limitations such as reliance on single data sources and disregard for crucial features like the trend in Arterial Blood Pressure (ABP) waveforms. To address these challenges, this paper proposes a novel multichannel deep learning framework that combines Convolutional Neural Networks (CNN) and Transformers for automatic IOH prediction. The model leverages four types of physiological waveforms, including ABP, ElectroCardioGram (ECG), photoplethysmography (PLE), and Carbon Dioxide (CO2), to capture both local and global information, enhancing information representation. Experimental results on retrospective data of 14,140 adult patients undergoing non-cardiac surgery from VitalDB, a public data repository of vital signs taken during surgeries in 10 operating rooms at Seoul National University Hospital (from January 6, 2005 to March 1, 2014), demonstrate that the proposed model consistently outperforms other methods, achieving superior AUROC values of 0.943, 0.928, and 0.923 at 5, 10, and 15 minutes before the event, respectively. These results demonstrate the powerfulness of multi-modal data source as well as the importance of ABP waveform trends. Additionally, the incorporation of multi-task learning and the attention mechanism validates the effectiveness and superiority of our model, highlighting its potential for proactive clinical interventions and improved postoperative patient outcomes.

Video visual relation inference aims at extracting the relation triplets in the form of$>$in videos. With the development of deep learning, existing approaches are designed based on data-driven neural networks. But the datasets are always biased in terms of objects and relation triplets, which make relation inference challenging. Existing approaches often describe the relationships from visual, spatial, and semantic characteristics. The semantic description plays a key role to indicate the potential linguistic connections between objects, that are crucial to transfer knowledge across relationships, especially for the determination of novel relations. However, in these works, the semantic features are not emphasized, but simply obtained by mapping object labels, which can not reflect sufficient linguistic meanings. To alleviate the above issues, we propose a novel network, termed Concept-Enhanced Relation Network (CERN), to facilitate video visual relation inference. Thanks to the attributes and linguistic contexts implied in concepts, the semantic representations aggregated with related concept knowledge of objects are of benefit to relation inference. To this end, we incorporate retrieved concepts with local semantics of objects via the gating mechanism to generate the concept-enhanced semantic representations. Extensive experimental results show that our approach has achieved state-of-the-art performance on two public datasets: ImageNet-VidVRD and VidOR.

Knowledge has long been cosnsidered a crucial part of natural language understanding. Many knowledge bases have been constructed, but none of them will ever be complete. Nevertheless, we argue that complete knowledge already exists in natural language. Most previous work on question answering retrieved such knowledge using traditional statistical methods, and consequently, the knowledge retrieved contained co-occurring phrases and could not provide guidance for model understanding and reasoning. Therefore, in addition to demonstrating the effectiveness of natural language knowledge in machine understanding, this study presents a novel knowledge retrieval approach that evaluates the importance of knowledge from a semantic perspective. Furthermore, we propose a knowledge revise mechanism that allows the model to revise the retrieved knowledge from local and global perspectives. We demonstrate our Semantic-rank-and-Knowledge-Revise-based Question Answering (SKR-QA) approach on two challenging multi-choice question and answering tasks: ARC–Challenge and OpenbookQA. Compared with the previous State-of-the-Art (SOTA) models, our work achieves consistent improvements. Moreover, the knowledge obtained by our method is more conducive to machine understanding, thus providing certain interpretability.

We focus on the task of reasoning over paragraph effects in situation, which requires a model to understand the cause and effect described in a background paragraph, and apply the knowledge to a novel situation.Existing works ignore the complicated reasoning process and solve it with a one-step "black box" model.Inspired by human cognitive processes, in this paper we propose a sequential approach for this task which explicitly models each step of the reasoning process with neural network modules.In particular, five reasoning modules are designed and learned in an end-to-end manner, which leads to a more interpretable model.Experimental results on the ROPES dataset demonstrate the effectiveness and explainability of our proposed approach.