Hongkai Zhao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Hongkai Zhao has a total of 30 co-authors. Here are the top co-authors by number of shared publications:
- Zhelong Wang
Key Laboratory of Intelligent Control and Optimization for Industrial Equipment of Ministry of Education, Dalian University of Technology, Dalian, China
6 shared publications - Hongyu Zhao
Key Laboratory of Intelligent Control and Optimization for Industrial Equipment of Ministry of Education, Dalian University of Technology, Dalian, China
5 shared publications - Sen Qiu
School of Control Science and Engineering, Dalian University of Technology, Dalian, China
4 shared publications - Ke Zhang
Dalian University of Technology
2 shared publications - Daoyong Peng
Central Hospital of Dalian University of Technology, Dalian, China
1 shared publication - Donghui Wu
School of Building Environment Engineering, Zhengzhou University of Light Industry, Zhengzhou, China
1 shared publication
Latest Publications
- 10 August 2026
- Published by Elsevier in Neurocomputing
- Vol. 703
Ada-FMP: An adaptive fluctuation-aware multi-scale patch network for time series forecasting
- 10 August 2026
- Published by Elsevier in Neurocomputing
- Vol. 703
Ada-FMP: An adaptive fluctuation-aware multi-scale patch network for time series forecasting
- 1 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. 73 (6) , 2227-2236
A Hierarchical Multimodal Framework for Sedation Monitoring in ICU Patients
In the intensive care unit (ICU), monitoring sedation levels is crucial. Clinicians often rely on intermittent behavioral scales like the Richmond Agitation-Sedation Scale (RASS), which can be subjective and delay timely interventions. While electroencephalogram (EEG) offers a continuous and non-invasive alternative, but the complexity of consciousness renders a unimodal signal insufficient for comprehensive representation. To address these challenges, we propose a novel multimodal deep learning framework, Hierarchical Multimodal Fusion with Dynamic Correction (HMDC), that synergistically integrates EEG with peripheral physiological signals including blood pressure, heart rate, and oxygen saturation. The architecture features a dual-stream pathway to process both raw temporal EEG data and its spectral features from spectrograms. These neural representations are then intelligently fused and refined by a Dynamic Correction Module using a confidence-weighting mechanism. The model was developed and validated on a dataset comprising 2,880 labeled RASS assessments from 105 ICU patients, with scores ranging from −5 (comatose) to +1 (restless). The HMDC framework achieved a classification accuracy of 83.8%, significantly outperforming unimodal and simpler fusion baselines. By providing a temporally precise and physiologically grounded sedation assessment, this integrative approach establishes a robust correlation between multimodal signal patterns and clinical states, offering clinicians a unified tool for optimizing sedative titration and potentially minimizing delirium risks.
- 1 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biomedical Engineering
- Vol. 73 (6) , 2227-2236
A Hierarchical Multimodal Framework for Sedation Monitoring in ICU Patients
In the intensive care unit (ICU), monitoring sedation levels is crucial. Clinicians often rely on intermittent behavioral scales like the Richmond Agitation-Sedation Scale (RASS), which can be subjective and delay timely interventions. While electroencephalogram (EEG) offers a continuous and non-invasive alternative, but the complexity of consciousness renders a unimodal signal insufficient for comprehensive representation. To address these challenges, we propose a novel multimodal deep learning framework, Hierarchical Multimodal Fusion with Dynamic Correction (HMDC), that synergistically integrates EEG with peripheral physiological signals including blood pressure, heart rate, and oxygen saturation. The architecture features a dual-stream pathway to process both raw temporal EEG data and its spectral features from spectrograms. These neural representations are then intelligently fused and refined by a Dynamic Correction Module using a confidence-weighting mechanism. The model was developed and validated on a dataset comprising 2,880 labeled RASS assessments from 105 ICU patients, with scores ranging from −5 (comatose) to +1 (restless). The HMDC framework achieved a classification accuracy of 83.8%, significantly outperforming unimodal and simpler fusion baselines. By providing a temporally precise and physiologically grounded sedation assessment, this integrative approach establishes a robust correlation between multimodal signal patterns and clinical states, offering clinicians a unified tool for optimizing sedative titration and potentially minimizing delirium risks.
- 30 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Artificial Intelligence
- Vol. PP (9) , 1-11
An End-to-End Dual-Attention Based Target Speaker Extraction Network for Explicit Optimization of Speaker Confusion
The task of target speaker extraction (TSE) involves isolating a specific speaker’s voice from a mixture of multiple audio sources. Most existing TSE methods primarily optimize the reconstruction quality of the extracted waveform while overlooking essential cues contained in the reference audio. Moreover, traditional TSE networks lack an interactive learning mechanism that jointly models mixed and reference embeddings, often leading to speaker confusion, particularly when there is a notable mismatch between the reference and target audio. To overcome these challenges, we propose a novel end-to-end dual-attention-based target speaker extraction (DATSE) network that explicitly addresses speaker confusion. DATSE employs an attention-weighted multi-scale speech encoder that adaptively emphasizes features at different temporal resolutions. In addition, it introduces an attention-based speaker aggregation module within the auxiliary speaker recognition network, which effectively integrates speaker-specific information and leverages mixed-audio cues to refine target speaker embeddings. Experimental results on the WSJ0-2mix-extr dataset show that, the proposed method not only achieves competitive performance across multiple metrics but also significantly reduces speaker confusion under various disturbances such as noise and reverberation, thereby effectively lowering the overall confusion rate.
- 30 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Artificial Intelligence
- Vol. PP (9) , 1-11
An End-to-End Dual-Attention Based Target Speaker Extraction Network for Explicit Optimization of Speaker Confusion
The task of target speaker extraction (TSE) involves isolating a specific speaker’s voice from a mixture of multiple audio sources. Most existing TSE methods primarily optimize the reconstruction quality of the extracted waveform while overlooking essential cues contained in the reference audio. Moreover, traditional TSE networks lack an interactive learning mechanism that jointly models mixed and reference embeddings, often leading to speaker confusion, particularly when there is a notable mismatch between the reference and target audio. To overcome these challenges, we propose a novel end-to-end dual-attention-based target speaker extraction (DATSE) network that explicitly addresses speaker confusion. DATSE employs an attention-weighted multi-scale speech encoder that adaptively emphasizes features at different temporal resolutions. In addition, it introduces an attention-based speaker aggregation module within the auxiliary speaker recognition network, which effectively integrates speaker-specific information and leverages mixed-audio cues to refine target speaker embeddings. Experimental results on the WSJ0-2mix-extr dataset show that, the proposed method not only achieves competitive performance across multiple metrics but also significantly reduces speaker confusion under various disturbances such as noise and reverberation, thereby effectively lowering the overall confusion rate.
- 20 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
ST-DECNet: A Spatiotemporal Graph Learning Framework for Epileptic Seizure Prediction via Dynamic Effective Connectivity in EEGs
Effective connectivity represents causal interactions in the brain, yet analyzing its nonlinear and dynamic properties in electroencephalogram (EEG) data, which is measured by recording electrical activity through scalp electrodes, remains a significant challenge. This work introduces a novel framework, spatio-temporal dynamic effective connectivity networks (ST-DECNet), designed to detect time-varying nonlinear effective connectivity networks (ECNs) in the brain. The approach begins with a gated temporal convolutional network based on neural Granger causality (GTCN-NGC) to estimate dynamic ECNs from multi-channel EEG signals, incorporating automatic lag selection. These ECNs are subsequently transformed into dynamic graph representations via a graph isomorphism network, where attention mechanisms capture intricate spatio-temporal dynamics. The effectiveness of ST-DECNet is demonstrated through its application to epileptic seizure prediction, revealing its capability to extract key biomarkers and enhance prediction accuracy. Experimental results highlight ST-DECNet’s potential for advancing automated seizure prediction and brain connectivity analysis.
- 20 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
ST-DECNet: A Spatiotemporal Graph Learning Framework for Epileptic Seizure Prediction via Dynamic Effective Connectivity in EEGs
Effective connectivity represents causal interactions in the brain, yet analyzing its nonlinear and dynamic properties in electroencephalogram (EEG) data, which is measured by recording electrical activity through scalp electrodes, remains a significant challenge. This work introduces a novel framework, spatio-temporal dynamic effective connectivity networks (ST-DECNet), designed to detect time-varying nonlinear effective connectivity networks (ECNs) in the brain. The approach begins with a gated temporal convolutional network based on neural Granger causality (GTCN-NGC) to estimate dynamic ECNs from multi-channel EEG signals, incorporating automatic lag selection. These ECNs are subsequently transformed into dynamic graph representations via a graph isomorphism network, where attention mechanisms capture intricate spatio-temporal dynamics. The effectiveness of ST-DECNet is demonstrated through its application to epileptic seizure prediction, revealing its capability to extract key biomarkers and enhance prediction accuracy. Experimental results highlight ST-DECNet’s potential for advancing automated seizure prediction and brain connectivity analysis.
- 20 March 2024
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 173, 108348
Toxicity prediction and classification of Gunqile-7 with small sample based on transfer learning method
- 20 March 2024
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 173, 108348
Toxicity prediction and classification of Gunqile-7 with small sample based on transfer learning method
- 27 September 2021
- Published by Wiley in International Journal of Intelligent Systems
- Vol. 37 (2) , 1646-1673
Sensor network oriented human motion capture via wearable intelligent system
- 27 September 2021
- Published by Wiley in International Journal of Intelligent Systems
- Vol. 37 (2) , 1646-1673
Sensor network oriented human motion capture via wearable intelligent system
- Open Access
- 1 June 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 15925-15934
A Dual Iterative Refinement Method for Non-rigid Shape Matching
In this work, a robust and efficient dual iterative refinement (DIR) method is proposed for dense correspondence between two nearly isometric shapes. The key idea is to use dual information, such as spatial and spectral, or local and global features, in a complementary and effective way, and extract more accurate information from current iteration to use for the next iteration. In each DIR iteration, starting from current correspondence, a zoom-in process at each point is used to select well matched anchor pairs by a local mapping distortion criterion. These selected anchor pairs are then used to align spectral features (or other appropriate global features) whose dimension adaptively matches the capacity of the selected anchor pairs. Thanks to the effective combination of complementary information in a data-adaptive way, DIR is not only efficient but also robust to render accurate results within a few iterations. By choosing appropriate dual features, DIR has the flexibility to handle patch and partial matching as well. Our comprehensive experiments on various data sets demonstrate the superiority of DIR over other state-of-the-art methods in terms of both accuracy and efficiency.
- 1 June 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 15925-15934
Open AccessA Dual Iterative Refinement Method for Non-rigid Shape Matching
In this work, a robust and efficient dual iterative refinement (DIR) method is proposed for dense correspondence between two nearly isometric shapes. The key idea is to use dual information, such as spatial and spectral, or local and global features, in a complementary and effective way, and extract more accurate information from current iteration to use for the next iteration. In each DIR iteration, starting from current correspondence, a zoom-in process at each point is used to select well matched anchor pairs by a local mapping distortion criterion. These selected anchor pairs are then used to align spectral features (or other appropriate global features) whose dimension adaptively matches the capacity of the selected anchor pairs. Thanks to the effective combination of complementary information in a data-adaptive way, DIR is not only efficient but also robust to render accurate results within a few iterations. By choosing appropriate dual features, DIR has the flexibility to handle patch and partial matching as well. Our comprehensive experiments on various data sets demonstrate the superiority of DIR over other state-of-the-art methods in terms of both accuracy and efficiency.