Selected work

Representative Papers

Leveraging Semantic Type Dependencies for Clinical Named Entity Recognition.

Mar 07, 2025AMIA ... Annual Symposium proceedings. AMIA Symposium

This paper addresses the underutilization of domain-specific semantic knowledge in clinical named entity recognition (NER). We propose a novel modeling approach that explicitly incorporates semantic type dependencies from the Unified Medical Language System (UMLS) to enhance contextual understanding. Specifically, we design a single-pass matrix encoding mechanism that represents multi-type semantic relations—e.g., “Disease–Symptom” or “Drug–Dosage”—between entity spans and context tokens as structured matrices, seamlessly integrated into a BiLSTM-GCN-CRF architecture. The method is compatible with clinical pretrained embeddings (e.g., BERT, BioBERT, UMLSBERT) without requiring additional fine-tuning. Evaluated on standard clinical benchmarks—including i2b2 and ShARe/CLEF—it achieves significant improvements in fine-grained NER performance, with average F1-score gains of 1.8–3.2 percentage points. Results demonstrate that explicit modeling of domain-specific semantic dependencies substantially enhances clinical NER accuracy and advances knowledge-guided NER paradigms.

1 citationsRead paper

A Deterministic Binary Fingerprinting Framework with Zero-Trained Feature Extraction for Sparse Count Matrices

Jul 16, 2026

This work addresses the high computational and memory costs of traditional methods—such as PCA or iterative optimization—for analyzing sparse count data from single-cell transcriptomics, k-mer spectra, and document term frequencies. The authors propose MMTB, a novel framework that deterministically maps samples to binary “thermometer” fingerprints via column-wise Min-Max normalization and a fixed threshold, eliminating the need for training, labels, or gradient-based optimization. To the best of our knowledge, this is the first method to achieve learning-free, deterministic binary encoding that preserves high correlation with L1 distance while enabling efficient Hamming-distance-based indexing. Evaluated on a mixture of three cell lines, MMTB achieves an NMI of 0.99 using only 188 bytes per cell—approximately a 10× reduction in memory compared to float32 embeddings—making it well-suited for resource-constrained, coarse-grained clustering tasks.

0 citationsRead paper

Gated Fusion Enhanced Multi-Scale Hierarchical Graph Convolutional Network for Stock Movement Prediction

Nov 03, 2025

Stock price forecasting is highly challenging due to market volatility and complex cross-stock dependencies. Existing multi-scale graph neural networks (GNNs) often neglect the influence of intra-stock attribute patterns on relational modeling and suffer from coarse-/fine-grained bias in multi-scale feature fusion. To address these issues, we propose a multi-scale hierarchical GNN framework. First, we design a dynamic graph construction mechanism that explicitly models time-varying topologies driven by intra-stock attribute patterns. Second, we introduce a top-down gated fusion strategy to jointly optimize coarse-grained global structural features and fine-grained local dynamic features. Our approach unifies spatiotemporal dependency modeling with multi-scale relational learning. Experiments on real-world stock data from both U.S. and Chinese markets demonstrate that our method achieves up to a 1.4% improvement in prediction accuracy and significantly outperforms mainstream baseline models in profit stability.

0 citationsRead paper

Robust High-Resolution Multi-Organ Diffusion MRI Using Synthetic-Data-Tuned Prompt Learning

Oct 17, 2025

Multi-shot whole-body diffusion-weighted imaging (DWI) suffers from severe motion artifacts due to respiration and peristalsis, while simultaneously achieving multi-organ coverage, multiple b-values, and high spatial resolution remains clinically challenging. Method: We propose LoSP-Prompt—a novel MRI reconstruction framework that introduces prompt learning into DWI reconstruction for the first time. It integrates high-order local smooth phase modeling with low-rank Hankel matrix optimization and enables end-to-end parameter learning on synthetic abdominal DWI data—without navigators or ground-truth annotations. Contribution/Results: LoSP-Prompt exhibits strong interpretability and cross-organ, cross-scanner generalizability. Validated on over 10,000 clinical images, it doubles spatial resolution, covers seven anatomical regions (e.g., liver, kidney, brain), and achieves radiologist-rated image quality scores of 4–5/5—significantly outperforming state-of-the-art methods.

0 citationsRead paper

DisMS-TS: Eliminating Redundant Multi-Scale Features for Time Series Classification

Jul 06, 2025

Existing multi-scale time series forecasting methods struggle to eliminate redundant shared features across scales, leading to imbalanced modeling of shared versus scale-specific features. To address this, we propose DisMS-TS—a novel framework that, for the first time, explicitly disentangles temporal representations into scale-shared and scale-specific components. DisMS-TS introduces a learnable temporal disentanglement module, coupled with two complementary regularizers: a consistency regularizer enforcing cross-scale invariance of shared features, and a discriminability regularizer enhancing the separability of scale-specific features. This end-to-end deep architecture jointly integrates multi-scale analysis and regularized feature disentanglement. Extensive experiments on multiple standard benchmarks demonstrate substantial improvements in time series classification performance, with up to a 9.71% absolute accuracy gain. The results validate both the effectiveness and generalizability of explicit disentangled representation learning for multi-scale time series classification.

0 citationsRead paper

Recent publications

Latest Papers

A Deterministic Binary Fingerprinting Framework with Zero-Trained Feature Extraction for Sparse Count Matrices

Jul 16, 2026

This work addresses the high computational and memory costs of traditional methods—such as PCA or iterative optimization—for analyzing sparse count data from single-cell transcriptomics, k-mer spectra, and document term frequencies. The authors propose MMTB, a novel framework that deterministically maps samples to binary “thermometer” fingerprints via column-wise Min-Max normalization and a fixed threshold, eliminating the need for training, labels, or gradient-based optimization. To the best of our knowledge, this is the first method to achieve learning-free, deterministic binary encoding that preserves high correlation with L1 distance while enabling efficient Hamming-distance-based indexing. Evaluated on a mixture of three cell lines, MMTB achieves an NMI of 0.99 using only 188 bytes per cell—approximately a 10× reduction in memory compared to float32 embeddings—making it well-suited for resource-constrained, coarse-grained clustering tasks.

0 citationsRead paper

Gated Fusion Enhanced Multi-Scale Hierarchical Graph Convolutional Network for Stock Movement Prediction

Nov 03, 2025

Stock price forecasting is highly challenging due to market volatility and complex cross-stock dependencies. Existing multi-scale graph neural networks (GNNs) often neglect the influence of intra-stock attribute patterns on relational modeling and suffer from coarse-/fine-grained bias in multi-scale feature fusion. To address these issues, we propose a multi-scale hierarchical GNN framework. First, we design a dynamic graph construction mechanism that explicitly models time-varying topologies driven by intra-stock attribute patterns. Second, we introduce a top-down gated fusion strategy to jointly optimize coarse-grained global structural features and fine-grained local dynamic features. Our approach unifies spatiotemporal dependency modeling with multi-scale relational learning. Experiments on real-world stock data from both U.S. and Chinese markets demonstrate that our method achieves up to a 1.4% improvement in prediction accuracy and significantly outperforms mainstream baseline models in profit stability.

0 citationsRead paper

Robust High-Resolution Multi-Organ Diffusion MRI Using Synthetic-Data-Tuned Prompt Learning

Oct 17, 2025

Multi-shot whole-body diffusion-weighted imaging (DWI) suffers from severe motion artifacts due to respiration and peristalsis, while simultaneously achieving multi-organ coverage, multiple b-values, and high spatial resolution remains clinically challenging. Method: We propose LoSP-Prompt—a novel MRI reconstruction framework that introduces prompt learning into DWI reconstruction for the first time. It integrates high-order local smooth phase modeling with low-rank Hankel matrix optimization and enables end-to-end parameter learning on synthetic abdominal DWI data—without navigators or ground-truth annotations. Contribution/Results: LoSP-Prompt exhibits strong interpretability and cross-organ, cross-scanner generalizability. Validated on over 10,000 clinical images, it doubles spatial resolution, covers seven anatomical regions (e.g., liver, kidney, brain), and achieves radiologist-rated image quality scores of 4–5/5—significantly outperforming state-of-the-art methods.

0 citationsRead paper

DisMS-TS: Eliminating Redundant Multi-Scale Features for Time Series Classification

Jul 06, 2025

Existing multi-scale time series forecasting methods struggle to eliminate redundant shared features across scales, leading to imbalanced modeling of shared versus scale-specific features. To address this, we propose DisMS-TS—a novel framework that, for the first time, explicitly disentangles temporal representations into scale-shared and scale-specific components. DisMS-TS introduces a learnable temporal disentanglement module, coupled with two complementary regularizers: a consistency regularizer enforcing cross-scale invariance of shared features, and a discriminability regularizer enhancing the separability of scale-specific features. This end-to-end deep architecture jointly integrates multi-scale analysis and regularized feature disentanglement. Extensive experiments on multiple standard benchmarks demonstrate substantial improvements in time series classification performance, with up to a 9.71% absolute accuracy gain. The results validate both the effectiveness and generalizability of explicit disentangled representation learning for multi-scale time series classification.

0 citationsRead paper

Transformer-Based Dual-Optical Attention Fusion Crowd Head Point Counting and Localization Network

May 11, 2025

To address low accuracy in crowd counting and head localization under dense occlusion and low-light conditions in drone-captured scenes, this paper proposes an end-to-end point detection framework leveraging visible-infrared dual-modal inputs. The method employs a Transformer-based backbone integrated with dual-stream CNN feature extraction, cross-modal attention, and explicit feature alignment mechanisms. We introduce two novel modules: the Dual-Optical Attention Fusion Module (DAFP) and the Adaptive Dual-Modal Feature Decomposition Fusion Module (AFDF), which jointly mitigate systematic cross-modal misalignment. Additionally, a spatial random shift data augmentation strategy is adopted to enhance model robustness. Evaluated on the DroneRGBT and GAIIC2 benchmarks under dense, low-light scenarios, our approach achieves an 18.7% reduction in Mean Absolute Error (MAE) and a 23.4% improvement in head localization accuracy over state-of-the-art methods.

0 citationsRead paper