Trong-Nghia Nguyen
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
Trong-Nghia Nguyen has a total of 43 co-authors. Here are the top co-authors by number of shared publications:
Bo-Gun Kho
6 shared publicationsSoo-Hyung Kim
6 shared publications- Soo-Hyung Kim
Chonnam National University
5 shared publications - Hyung-Jeong Yang
Chonnam National University
4 shared publications Guee-Sang Lee
3 shared publicationsHyung-Jeong Yang
3 shared publications
Latest Publications
- Open Access
- 11 November 2026
- Published by MDPI AG in Healthcare
- Vol. 14 (22) , 3869
Multi-Type Clinical Interaction Modeling for Identifying Factors Associated with Cardiovascular Disease in Obstructive Sleep Apnea via an Interpretable Heterogeneous Graph Transformer
Background/Objectives: Obstructive sleep apnea (OSA) is a prevalent sleep disorder strongly associated with an increased risk of cardiovascular disease (CVD). The relationship between OSA and CVD is complex, involving multiple physiological and clinical factors. Identifying key contributors to cardiovascular risk in OSA patients is essential for early intervention, personalized treatment, and improved patient outcomes. Methods: We propose a heterogeneous graph learning model, RiskHGT, to identify influential clinical factors and interaction patterns associated with CVD in a cross-sectional analysis distinguishing OSA patients with and without prevalent CVD. By leveraging the advantages of heterogeneous graphs in modeling multi-relational medical data, RiskHGT captures intricate relationships among clinical factors and enables interpretation of the most influential factors associated with CVD in OSA patients. The model was trained and evaluated on two large-scale clinical sleep cohorts, the Sleep Heart Health Study (SHHS) and the Multi-Ethnic Study of Atherosclerosis (MESA), with predictive performance assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision–recall curve (AUPRC). Results: RiskHGT achieved superior predictive performance compared with baseline machine learning and graph-based models, reaching an AUROC of up to 0.94 and an AUPRC of up to 0.90. Fasting glucose, diastolic blood pressure, smoking status, and age were identified as the factors most strongly associated with CVD in OSA patients. Conclusions: RiskHGT provides interpretable, attention-based insights into clinical factors associated with CVD in patients with OSA. The framework may have potential for clinical use, but external and prospective validation is required before any clinical application.
- 11 November 2026
- Published by MDPI AG in Healthcare
- Vol. 14 (22) , 3869
Open AccessMulti-Type Clinical Interaction Modeling for Identifying Factors Associated with Cardiovascular Disease in Obstructive Sleep Apnea via an Interpretable Heterogeneous Graph Transformer
Background/Objectives: Obstructive sleep apnea (OSA) is a prevalent sleep disorder strongly associated with an increased risk of cardiovascular disease (CVD). The relationship between OSA and CVD is complex, involving multiple physiological and clinical factors. Identifying key contributors to cardiovascular risk in OSA patients is essential for early intervention, personalized treatment, and improved patient outcomes. Methods: We propose a heterogeneous graph learning model, RiskHGT, to identify influential clinical factors and interaction patterns associated with CVD in a cross-sectional analysis distinguishing OSA patients with and without prevalent CVD. By leveraging the advantages of heterogeneous graphs in modeling multi-relational medical data, RiskHGT captures intricate relationships among clinical factors and enables interpretation of the most influential factors associated with CVD in OSA patients. The model was trained and evaluated on two large-scale clinical sleep cohorts, the Sleep Heart Health Study (SHHS) and the Multi-Ethnic Study of Atherosclerosis (MESA), with predictive performance assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision–recall curve (AUPRC). Results: RiskHGT achieved superior predictive performance compared with baseline machine learning and graph-based models, reaching an AUROC of up to 0.94 and an AUPRC of up to 0.90. Fasting glucose, diastolic blood pressure, smoking status, and age were identified as the factors most strongly associated with CVD in OSA patients. Conclusions: RiskHGT provides interpretable, attention-based insights into clinical factors associated with CVD in patients with OSA. The framework may have potential for clinical use, but external and prospective validation is required before any clinical application.
- Open Access
- 24 August 2026
- Published by Elsevier in SSRN Electronic Journal
A Cross-Attention Survival Network for Early Warning of Acute Clinical Deterioration from Vital-Sign Trajectories
Background. Acute clinical deterioration in hospitalised patients is a leading cause of preventable in-hospital adverse outcomes. Current early-warning scores depend on static admission-time vital signs and miss the prognostic signal embedded in longitudinal trajectories, while existing machine-learning models require institution-specific feature engineering that limits cross-site applicability.Methods. We conducted a retrospective cohort study at Chonnam National University Hospital (CNUH) comprising 84,049 hourly vital-sign records across 28 features from a Korean general-ward cohort (41 critical events; 4.7 % event rate). We propose TraceSURV (Temporal Representation Attentive Cross-Encoders for Survival Prediction), an end-to-end deep survival network that operates directly on raw multivariate vital-sign sequences without an external feature-engineering pipeline, jointly modeling temporal dynamics and inter-feature interactions via cross-attention. TraceSURV was benchmarked against 18 comparators spanning six methodological families under five-fold cross-validation, and evaluated for architectural generalizability on PhysioNet 2019 Set A (N=20,029) with all models independently retrained on that cohort (not cross-site model transfer).Results. TraceSURV achieved the best tdAUC (0.994 [0.982–0.999]) and best calibration (IBS 0.008 [0.005–0.012]) among all 19 models, and its C-index (0.969 [0.935–0.991]) substantially outperformed clinical early-warning scores (MEWS C = 0.744; NEWS C = 0.741). Given the small primary cohort (41 events), C-index confidence intervals are wide and rankings of the four Tier-1 models overlap (all pairwise p>0.05); TraceSURV's statistically distinguishable advantages were in tdAUC (0.994 vs. 0.980–0.990 for peers) and calibration (IBS 0.008 vs. 0.014–0.022 for peers). In the architectural generalisability experiment, TraceSURV achieved the highest C-index when retrained on PhysioNet (0.916 [0.907–0.923]) with the smallest performance drop (ΔC=−0.053). Attention weights identified Sodium, Platelet, Lactate, and Potassium as the dominant prognostic features, consistent with multi-organ failure pathophysiology; these patterns are hypothesis-generating rather than definitive causal attributions.Conclusion. End-to-end temporal learning on raw vital-sign sequences achieves superior calibration and robust cross-site architectural generalisability, with intrinsic cross-attention weights providing a native clinical interpretation pathway. TraceSURV offers a practical, pipeline-free route to early deterioration detection across heterogeneous clinical environments.
- 24 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Cross-Attention Survival Network for Early Warning of Acute Clinical Deterioration from Vital-Sign Trajectories
Background. Acute clinical deterioration in hospitalised patients is a leading cause of preventable in-hospital adverse outcomes. Current early-warning scores depend on static admission-time vital signs and miss the prognostic signal embedded in longitudinal trajectories, while existing machine-learning models require institution-specific feature engineering that limits cross-site applicability.Methods. We conducted a retrospective cohort study at Chonnam National University Hospital (CNUH) comprising 84,049 hourly vital-sign records across 28 features from a Korean general-ward cohort (41 critical events; 4.7 % event rate). We propose TraceSURV (Temporal Representation Attentive Cross-Encoders for Survival Prediction), an end-to-end deep survival network that operates directly on raw multivariate vital-sign sequences without an external feature-engineering pipeline, jointly modeling temporal dynamics and inter-feature interactions via cross-attention. TraceSURV was benchmarked against 18 comparators spanning six methodological families under five-fold cross-validation, and evaluated for architectural generalizability on PhysioNet 2019 Set A (N=20,029) with all models independently retrained on that cohort (not cross-site model transfer).Results. TraceSURV achieved the best tdAUC (0.994 [0.982–0.999]) and best calibration (IBS 0.008 [0.005–0.012]) among all 19 models, and its C-index (0.969 [0.935–0.991]) substantially outperformed clinical early-warning scores (MEWS C = 0.744; NEWS C = 0.741). Given the small primary cohort (41 events), C-index confidence intervals are wide and rankings of the four Tier-1 models overlap (all pairwise p>0.05); TraceSURV's statistically distinguishable advantages were in tdAUC (0.994 vs. 0.980–0.990 for peers) and calibration (IBS 0.008 vs. 0.014–0.022 for peers). In the architectural generalisability experiment, TraceSURV achieved the highest C-index when retrained on PhysioNet (0.916 [0.907–0.923]) with the smallest performance drop (ΔC=−0.053). Attention weights identified Sodium, Platelet, Lactate, and Potassium as the dominant prognostic features, consistent with multi-organ failure pathophysiology; these patterns are hypothesis-generating rather than definitive causal attributions.Conclusion. End-to-end temporal learning on raw vital-sign sequences achieves superior calibration and robust cross-site architectural generalisability, with intrinsic cross-attention weights providing a native clinical interpretation pathway. TraceSURV offers a practical, pipeline-free route to early deterioration detection across heterogeneous clinical environments.
- 20 February 2026
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 119
Band2CleanFormer: Integrating band-specific processing and inter-band attention for robust EEG denoising
- 20 February 2026
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 119
Band2CleanFormer: Integrating band-specific processing and inter-band attention for robust EEG denoising
- 22 November 2025
- Published by Springer Nature
- p. 115-126
MediFusion-Flex: An Adaptive Multimodal Deep Learning Framework for Clinical Deterioration Prediction in Emergency Medicine
- 22 November 2025
- Published by Springer Nature
- p. 115-126
MediFusion-Flex: An Adaptive Multimodal Deep Learning Framework for Clinical Deterioration Prediction in Emergency Medicine
- Open Access
- 3 September 2025
- Published by PeerJ in PeerJ Computer Science
- Vol. 11, e3158
Dual-stream transformer approach for pain assessment using visual-physiological data modeling
Automatic pain assessment involves accurately recognizing and quantifying pain, dependent on the data modality that may originate from various sources such as video and physiological signals. Traditional pain assessment methods rely on subjective self-reporting, which limits their objectivity, consistency, and overall effectiveness in clinical settings. While machine learning offers a promising alternative, many existing approaches rely on a single data modality, which may not adequately capture the multifaceted nature of pain-related responses. In contrast, multimodal approaches can provide a more comprehensive understanding by integrating diverse sources of information. To address this, we propose a dual-stream framework for classifying physiological and behavioral correlates of pain that leverages multimodal data to enhance robustness and adaptability across diverse clinical scenarios. Our framework begins with masked autoencoder pre-training for each modality: facial video and multivariate bio-psychological signals, to compress the raw temporal input into meaningful representations, enhancing their ability to capture complex patterns in high-dimensional data. In the second stage, the complete classifier consists of a dual hybrid positional encoding embedding and cross-attention fusion. The pain assessment evaluations reveal our model’s superior performance on the AI4Pain and BioVid datasets for electrode-based and heat-induced settings.
- 3 September 2025
- Published by PeerJ in PeerJ Computer Science
- Vol. 11, e3158
Open AccessDual-stream transformer approach for pain assessment using visual-physiological data modeling
Automatic pain assessment involves accurately recognizing and quantifying pain, dependent on the data modality that may originate from various sources such as video and physiological signals. Traditional pain assessment methods rely on subjective self-reporting, which limits their objectivity, consistency, and overall effectiveness in clinical settings. While machine learning offers a promising alternative, many existing approaches rely on a single data modality, which may not adequately capture the multifaceted nature of pain-related responses. In contrast, multimodal approaches can provide a more comprehensive understanding by integrating diverse sources of information. To address this, we propose a dual-stream framework for classifying physiological and behavioral correlates of pain that leverages multimodal data to enhance robustness and adaptability across diverse clinical scenarios. Our framework begins with masked autoencoder pre-training for each modality: facial video and multivariate bio-psychological signals, to compress the raw temporal input into meaningful representations, enhancing their ability to capture complex patterns in high-dimensional data. In the second stage, the complete classifier consists of a dual hybrid positional encoding embedding and cross-attention fusion. The pain assessment evaluations reveal our model’s superior performance on the AI4Pain and BioVid datasets for electrode-based and heat-induced settings.
- 7 October 2024
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 100
Temporal variational autoencoder model for in-hospital clinical emergency prediction
- 7 October 2024
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 100
Temporal variational autoencoder model for in-hospital clinical emergency prediction
- Open Access
- 23 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 117176-117202
Explainable Deep Contrastive Federated Learning System for Early Prediction of Clinical Status in Intensive Care Unit
Early identification of patients’ clinical status plays a critical role in intensive care unit (ICU) care. The increased adoption of electronic health records (EHRs) in the ICU creates prospects for deep learning (DL) application systems in this discipline. However, monitoring and prediction systems in the ICU encounter problems with security, alarm errors, and interpretation. This research presents deep contrastive federated learning (Deep-CFL), an approach that leverages explainable AI (XAI), CFL, and imbalanced supervised learning techniques to address these problems. CFL introduces an innovative approach to minimize the difference in local and global model prediction ability while increasing the gap in prediction performance of the current local model and its previous model in a communication round. When paired with imbalanced learning, this strategy substantially mitigates error alarm problems while ensuring data security. The XAI technique, specifically integrated gradient, is employed to refine the DL-based model architecture to enhance system interpretability. Extensive experiments and in-depth analyses across three significant clinical datasets highlight the superior performance of Deep-CFL over local and centralized learning-based approaches. The results involving 25, 329 patients admitted to Chonnam National University Hospital reveal that Deep-CFL, with an area under the receiver operating characteristics curve of 0.879, an area under the precision-recall curve of 0.886, and an average precision of 0.884, surpasses systems based on centralized learning while reducing the late alarm rate by up to 10.3%.
- 23 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 117176-117202
Open AccessExplainable Deep Contrastive Federated Learning System for Early Prediction of Clinical Status in Intensive Care Unit
Early identification of patients’ clinical status plays a critical role in intensive care unit (ICU) care. The increased adoption of electronic health records (EHRs) in the ICU creates prospects for deep learning (DL) application systems in this discipline. However, monitoring and prediction systems in the ICU encounter problems with security, alarm errors, and interpretation. This research presents deep contrastive federated learning (Deep-CFL), an approach that leverages explainable AI (XAI), CFL, and imbalanced supervised learning techniques to address these problems. CFL introduces an innovative approach to minimize the difference in local and global model prediction ability while increasing the gap in prediction performance of the current local model and its previous model in a communication round. When paired with imbalanced learning, this strategy substantially mitigates error alarm problems while ensuring data security. The XAI technique, specifically integrated gradient, is employed to refine the DL-based model architecture to enhance system interpretability. Extensive experiments and in-depth analyses across three significant clinical datasets highlight the superior performance of Deep-CFL over local and centralized learning-based approaches. The results involving 25, 329 patients admitted to Chonnam National University Hospital reveal that Deep-CFL, with an area under the receiver operating characteristics curve of 0.879, an area under the precision-recall curve of 0.886, and an average precision of 0.884, surpasses systems based on centralized learning while reducing the late alarm rate by up to 10.3%.
- 17 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Intelligent Systems
- Vol. 39 (6) , 58-69
Multigradient Siamese Temporal Model for the Prediction of Clinical Events in Rapid Response Systems
Early identification of emergency situations for intensive care unit (ICU) patients is a critical component of precision medicine. We present a Multi-gradient Siamese Temporal (MG-SiameseTS) model to predict patients’ clinical status within an 8-h window and trigger an alarm when anticipating an emergent situation. It incorporates gradient extractors, contrastive learning, focal loss, and imbalanced loss to address challenges related to bias, data quality, and interpretability. Experimental results using a practical dataset collected from the rapid response team at Chonnam National University Hospital in Korea demonstrate that the MG-SiameseTS model exhibits superior predictive capabilities for clinical events compared to several state-of-the-art methods.
- 17 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Intelligent Systems
- Vol. 39 (6) , 58-69
Multigradient Siamese Temporal Model for the Prediction of Clinical Events in Rapid Response Systems
Early identification of emergency situations for intensive care unit (ICU) patients is a critical component of precision medicine. We present a Multi-gradient Siamese Temporal (MG-SiameseTS) model to predict patients’ clinical status within an 8-h window and trigger an alarm when anticipating an emergent situation. It incorporates gradient extractors, contrastive learning, focal loss, and imbalanced loss to address challenges related to bias, data quality, and interpretability. Experimental results using a practical dataset collected from the rapid response team at Chonnam National University Hospital in Korea demonstrate that the MG-SiameseTS model exhibits superior predictive capabilities for clinical events compared to several state-of-the-art methods.
- 1 June 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 5770-5776
Ensemble Spatial and Temporal Vision Transformer for Action Units Detection
Facial Action Units detection (FAUs) represents a fine-grained classification problem that involves identifying different units on the human face, as defined by the Facial Action Coding System. In this paper, we present a simple yet efficient Vision Transformer-based approach for addressing the task of Action Units (AU) detection in the context of Affective Behavior Analysis in-the-wild (ABAW) competition. We employ the Video Vision Transformer(ViViT) Network to capture the temporal facial change in the video. Besides, to reduce massive size of the Vision Transformers model, we replace the ViViT feature extraction layers with the CNN backbone (Regnet). Our model outperform the baseline model of ABAW 2023 challenge [8], with a notable 14% difference in result. Our team has achieved a position within the top 5 teams in the ABAW 2023 competition, scoring slightly below the top three and four teams by a narrow margin of 0.27% and 0.43%, respectively.
- 1 June 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 5770-5776
Ensemble Spatial and Temporal Vision Transformer for Action Units Detection
Facial Action Units detection (FAUs) represents a fine-grained classification problem that involves identifying different units on the human face, as defined by the Facial Action Coding System. In this paper, we present a simple yet efficient Vision Transformer-based approach for addressing the task of Action Units (AU) detection in the context of Affective Behavior Analysis in-the-wild (ABAW) competition. We employ the Video Vision Transformer(ViViT) Network to capture the temporal facial change in the video. Besides, to reduce massive size of the Vision Transformers model, we replace the ViViT feature extraction layers with the CNN backbone (Regnet). Our model outperform the baseline model of ABAW 2023 challenge [8], with a notable 14% difference in result. Our team has achieved a position within the top 5 teams in the ABAW 2023 competition, scoring slightly below the top three and four teams by a narrow margin of 0.27% and 0.43%, respectively.