Dawen Sun

Dr. Dawen Sun


Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 December 2024 in IEEE Transactions on Neural Systems and Rehabilitation Engineering
Verified
Journal Article
Open access
ZKZhejun Kuang
Zhejun Kuang
JWJingrui Wang
Jingrui Wang
DSDawen Sun
Dawen Sun
JZJian Zhao
Jian Zhao
LSLijuan Shi
Lijuan Shi
YZYusheng Zhu
Yusheng Zhu
Rehabilitation training is essential for the recovery of patients with conditions such as stroke and Parkinson’s disease. However, traditional skeletal-based assessments often fail to capture the subtle movement qualities necessary for personalized care and are not optimized for scoring tasks. To address these limitations, we propose a hierarchical contrastive learning framework that integrates multi-view skeletal data, combining both positional and angular joint information. This integration enhances the framework’s ability to detect subtle variations in movement during rehabilitation exercises. In addition, we introduce a novel contrastive loss function specifically designed for regression tasks. This new approach yields substantial improvements over existing state-of-the-art models, achieving over a 30% reduction in mean absolute deviation on both the KIMORE and UIPRMD datasets. The framework demonstrates robustness in capturing both global and local movement characteristics, which are critical for accurate clinical evaluations. By precisely quantifying action quality, the framework supports the development of more targeted, personalized rehabilitation plans and shows strong potential for broad application in rehabilitation practices as well as in a wider range of motion assessment tasks.

23 September 2024 in Sensors
Verified
Journal Article
Open access
PDF file
ZKZhejun Kuang
Zhejun Kuang
XXXingbo Xiong
Xingbo Xiong
GWGang Wu
Gang Wu
FWFeng Wang
Feng Wang
JZJian Zhao
Jian Zhao
DSDawen Sun
Dawen Sun
Trigger–action programming (TAP) enables users to automate Internet of Things (IoT) devices by creating rules such as “IF Device1.TriggerState is triggered, THEN Device2.ActionState is executed”. As the number of IoT devices grows, the combination space between the functions provided by devices expands, making manual rule creation time-consuming for end-users. Existing TAP recommendation systems enhance the efficiency of rule discovery but face two primary issues: they ignore the association of rules between users and fail to model collaborative information among users. To address these issues, this article proposes a graph contrastive learning-based recommendation system for TAP rules, named GCL4TAP. In GCL4TAP, we first devise a data partitioning method called DATA2DIV, which establishes cross-user rule relationships and is represented by a user–rule bipartite graph. Then, we design a user–user graph to model the similarities among users based on the categories and quantities of devices that they own. Finally, these graphs are converted into low-dimensional vector representations of users and rules using graph contrastive learning techniques. Extensive experiments conducted on a real-world smart home dataset demonstrate the superior performance of GCL4TAP compared to other state-of-the-art methods.

1 January 2026 in Applied Sciences
Verified
Journal Article
Open access
PDF file
ZKZhejun Kuang
Zhejun Kuang
HGHaobo Gao
Haobo Gao
JYJiaxuan Yu
Jiaxuan Yu
DSDawen Sun
Dawen Sun
JZJian Zhao
Jian Zhao
LSLei Sun
Lei Sun
With growing public health awareness, accurate food nutrition estimation plays an increasingly important role in dietary management and disease prevention. The main bottleneck lies in how to effectively integrate multi-source heterogeneous information. We propose RDINet, a multimodal network that fuses RGB appearance, depth geometry, and ingredient semantics for food nutrition estimation. It comprises two core modules: The RGB-D fusion module integrates the textural appearance of RGB images and the 3D shape information conveyed by depth images through a channel–spatial attention mechanism, achieving a joint understanding of food appearance and geometric morphology without explicit 3D reconstruction; the ingredient fusion module embeds ingredient information into visual features via attention mechanisms, enabling the model to fully leverage components that are visually difficult to discern or prone to confusion, thereby activating corresponding nutritional reasoning pathways and achieving cross-modal inference from explicit observations to latent attributes. Experimental results on the Nutrition5k dataset show that RDINet achieves percentage mean absolute errors (PMAE) of 14.9%, 11.2%, 19.7%, 18.9%, and 19.5% for estimating calories, mass, fat, carbohydrates, and protein, respectively, with a mean PMAE of 16.8% across all metrics, outperforming existing mainstream methods. The results demonstrate that the appearance–geometry–semantics fusion framework is effective.

27 September 2024 in Journal of Computational Design and Engineering
Verified
Journal Article
Open access
PDF file
ZKZhejun Kuang
Zhejun Kuang
JWJingrui Wang
Jingrui Wang
DSDawen Sun
Dawen Sun
JZJian Zhao
Jian Zhao
LSLijuan Shi
Lijuan Shi
XXXingbo Xiong
Xingbo Xiong
In numerous academic and practical settings, it is common to progressively add data features to existing models as data evolves due to new discoveries and ongoing data collection efforts. Incremental attribute learning focuses on incorporating these new feature dimensions efficiently without retraining the entire model. The key challenge for incremental attribute learning is adapting models to include new attributes dynamically while maintaining their performance on existing data formats. This research leverages strategies from existing incremental learning practices, applying them to incremental attribute learning through a new framework that utilizes knowledge distillation to prevent catastrophic forgetting. Additionally, this study proposes an embedding layer algorithm to manage the variability of input features in incremental learning effectively. Extensive experiments across 11 tabular datasets, covering binary classification, multi-class classification, and regression tasks, indicate that our approach surpasses traditional incremental attribute learning methods and conventional retraining in terms of accuracy and efficiency. Compared to existing incremental attribute learning models, our model requires only a fraction of the training time and achieves a relative improvement of 7.97% in classification tasks. Furthermore, our approach exhibits minimal forgetfulness of old format data, with a forgetting rate of approximately 1% per increment stage, while maintaining strong performance on new format data. Despite its advantages, our method is limited to tabular data and requires additional storage space. We encourage future research to delve deeper into incremental attribute learning and its broader applications, aiming to consistently enhance model performance as new attributes are introduced. Additionally, we suggest exploring methods for refining and selecting soft labels or intermediate vectors to optimize storage efficiency while effectively leveraging the original model’s knowledge.

9 February 2026 in Frontiers in Medicine
Verified
Journal Article
Open access
ZKZhejun Kuang
Zhejun Kuang
XYXingxue Yan
Xingxue Yan
JYJiaxuan Yu
Jiaxuan Yu
DSDawen Sun
Dawen Sun
JZJian Zhao
Jian Zhao
LSLei Sun
Lei Sun
Introduction: Brain stroke is one of the leading causes of death and disability worldwide, and accurate lesion segmentation from MRI is critical for clinical diagnosis and treatment planning. However, existing methods struggle with the high variability of stroke lesions in size and morphology. In particular, they fail to detect small lesions due to the limited receptive fields of CNNs and the computational inefficiency of Transformer-based approaches. To address these challenges, we propose LBMNet, a novel CNN–Mamba network that integrates multi-scale convolutional encoding with Mamba-based decoding. Methods: Owing to the high heterogeneity of stroke lesions, the encoder design employs a top-down LSC module to capture cross-scale representations. The decoder designs the BSC-Mamba (Bidirectional Spatial Context Mamba) model, integrating bidirectional state space modeling with adaptive spatial convolutions to enhance local feature information while modeling global dependencies with linear complexity. Furthermore, asymmetric adaptive gated feature fusion (BAGF) bridges the semantic gap by selectively merging encoder and decoder features, suppressing redundant information whilst highlighting critical lesion details. Results: Extensive experiments on two benchmark datasets demonstrate state-of-the-art performance, achieving Dice coefficients of 67.57% on ATLAS v2.0 and 82.03% on ISLES 2022. Compared with existing CNN, Transformer, and hybrid models, LBMNet shows significant improvements in small lesion segmentation. This study presents a robust and efficient framework with strong clinical potential for accurate stroke lesion segmentation across diverse lesion sizes and morphologies.

23 October 2026 in Brain Sciences
Verified
Journal Article
Open access
PDF file
ZKZhejun Kuang
Zhejun Kuang
JWJunyi Wu
Junyi Wu
DSDawen Sun
Dawen Sun
JZJian Zhao
Jian Zhao
YWYifan Wang
Yifan Wang
Background: Patient-specific seizure prediction remains challenging because preictal EEG patterns and channel relevance vary substantially across individuals. Methods: We propose a CNN-GCN-Diffusion Network (CGD-Net) integrating channel-wise temporal encoding, montage-informed graph modeling, bag-level aggregation, and diffusion-prior-guided soft channel gating. The patient-specific channel prior was estimated exclusively from training data. CGD-Net was evaluated on 19 CHB-MIT recording cases using seizure-group-aware chronological training, validation, and test partitions. Robustness was assessed using random seeds 42, 52, and 62. Results: Across the three seeds, mean accuracy ranged from 74.84% to 75.87%, and mean AUC ranged from 81.82% to 82.01%. In the seed-42 component comparison, the full model achieved the highest mean accuracy, AUC, F1-score, and specificity. Recording-case-level Wilcoxon tests with Holm correction showed significant improvements in accuracy, AUC, and F1-score over the No bag, No diffusion, and No prior configurations, whereas differences from Random graph were not significant after correction. Channel-importance patterns were highly stable across seeds, with a mean pairwise Spearman correlation of 0.9819 and a mean Top-5 Jaccard index of 0.8730, while hub importance showed moderate patient-dependent stability. Conclusions: These results support the combined use of temporal aggregation, montage-informed graph modeling, and prior-guided channel modulation for retrospective preictal-versus-interictal classification. The explanations represent model-level attributions rather than clinically validated seizure-onset regions. Event-level sensitivity and false-alarm rates were not evaluated, and continuous-recording validation is required before clinical warning-system deployment.