Won-Seok Kang
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
Won-Seok Kang has a total of 34 co-authors. Here are the top co-authors by number of shared publications:
- Chang-Sik Son
Division of Intelligent Robot, Daegu Gyeongbuk Institute of Science and Technology (DGIST), Daegu 42988, Republic of Korea
7 shared publications - Cheil Moon
Department of Brain Sciences, Graduate School, Daegu Gyeongbuk Institute of Science and Technology (DGIST), Daegu, Republic of Korea
2 shared publications - Byungrak Son
DGIST
1 shared publication Chan-Beom Kim
1 shared publication- Dong-Ha Lee1 shared publication
- Dong-Ha Lee
Division of Intelligent Robot, DGIST, Daegu 42988, Korea
1 shared publication
Latest Publications
- Open Access
- 19 August 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (16) , 9134
Depthwise-Separable U-Net for Wearable Sensor-Based Human Activity Recognition
In wearable sensor-based human activity recognition (HAR), the traditional sliding window method encounters the challenge of multiclass windows in which multiple actions are combined within a single window. To address this problem, an approach that predicts activities at each point in time within a sequence has been proposed, and U-Net-based models have proven to be effective owing to their excellent space-time feature restoration capabilities. However, these models have limitations in that they are prone to overfitting owing to their large number of parameters and are not suitable for deployment. In this study, a lightweight U-Net was designed by replacing all standard U-Net convolutions with depthwise separable convolutions to implement dense prediction. Compared with existing U-Net-based models, the proposed model reduces the number of parameters by 57–89%. When evaluated on three benchmark datasets (MHEALTH, PAMAP2, and WISDM) using subject-independent splits, the performance of the proposed model was equal to or superior to that of all comparison models. Notably, on the MHEALTH dataset, which was collected in an uncontrolled environment, the proposed model improved accuracy by 7.89%, demonstrating its applicability to real-world wearable HAR systems.
- 19 August 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (16) , 9134
Open AccessDepthwise-Separable U-Net for Wearable Sensor-Based Human Activity Recognition
In wearable sensor-based human activity recognition (HAR), the traditional sliding window method encounters the challenge of multiclass windows in which multiple actions are combined within a single window. To address this problem, an approach that predicts activities at each point in time within a sequence has been proposed, and U-Net-based models have proven to be effective owing to their excellent space-time feature restoration capabilities. However, these models have limitations in that they are prone to overfitting owing to their large number of parameters and are not suitable for deployment. In this study, a lightweight U-Net was designed by replacing all standard U-Net convolutions with depthwise separable convolutions to implement dense prediction. Compared with existing U-Net-based models, the proposed model reduces the number of parameters by 57–89%. When evaluated on three benchmark datasets (MHEALTH, PAMAP2, and WISDM) using subject-independent splits, the performance of the proposed model was equal to or superior to that of all comparison models. Notably, on the MHEALTH dataset, which was collected in an uncontrolled environment, the proposed model improved accuracy by 7.89%, demonstrating its applicability to real-world wearable HAR systems.
- Open Access
- 18 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 128802-128816
Temporal and Modality Awareness-Based Lightweight Residual Network With Attention Mechanism for Human Activity Recognition Using a Lower-Limb Exoskeleton Robot
Although many human activity recognition (HAR) models have achieved high accuracy, their computational complexity often limits deployment in systems with constrained hardware resources, such as wearable lower-limb exoskeletons. In addition, existing models frequently overlook the complementary nature of multimodal sensor signals, focusing primarily on temporal dynamics while underutilizing modality-specific information. To address these issues, this study proposes a lightweight residual network for recognizing diverse locomotion activities across varying terrains using multimodal sensing data. The model adopts an asymmetric convolutional architecture composed of depthwise and pointwise layers to efficiently capture temporal and modality-specific features while significantly reducing the number of trainable parameters. A channel-attention block is further integrated to emphasize salient fused features. Evaluations on the Walking Assist Wearable Robot Motion dataset, which contains kinematic and postural signals from 500 adults using a lower-limb exoskeleton, demonstrated that the proposed model achieves an accuracy of 98.23% and a macro F1 score of 98.21%, with only 48,037 parameters. This outperforms four hybrid deep learning baselines while reducing the parameter count by 5.5–12.9 times. To assess generalizability, additional experiments were conducted on four benchmark datasets—UCI-HAR, HAPT, PAMAP2, and WISDM—under varying batch size conditions. The proposed model consistently achieved competitive or superior macro F1 scores of $0.9627~\pm ~0.0046$ , $0.776~\pm ~0.0138$ , $0.8937~\pm ~0.0121$ , and $0.9736~\pm ~0.0025$ , respectively, confirming its robustness and adaptability across diverse real-world HAR scenarios.
- 18 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 128802-128816
Open AccessTemporal and Modality Awareness-Based Lightweight Residual Network With Attention Mechanism for Human Activity Recognition Using a Lower-Limb Exoskeleton Robot
Although many human activity recognition (HAR) models have achieved high accuracy, their computational complexity often limits deployment in systems with constrained hardware resources, such as wearable lower-limb exoskeletons. In addition, existing models frequently overlook the complementary nature of multimodal sensor signals, focusing primarily on temporal dynamics while underutilizing modality-specific information. To address these issues, this study proposes a lightweight residual network for recognizing diverse locomotion activities across varying terrains using multimodal sensing data. The model adopts an asymmetric convolutional architecture composed of depthwise and pointwise layers to efficiently capture temporal and modality-specific features while significantly reducing the number of trainable parameters. A channel-attention block is further integrated to emphasize salient fused features. Evaluations on the Walking Assist Wearable Robot Motion dataset, which contains kinematic and postural signals from 500 adults using a lower-limb exoskeleton, demonstrated that the proposed model achieves an accuracy of 98.23% and a macro F1 score of 98.21%, with only 48,037 parameters. This outperforms four hybrid deep learning baselines while reducing the parameter count by 5.5–12.9 times. To assess generalizability, additional experiments were conducted on four benchmark datasets—UCI-HAR, HAPT, PAMAP2, and WISDM—under varying batch size conditions. The proposed model consistently achieved competitive or superior macro F1 scores of $0.9627~\pm ~0.0046$ , $0.776~\pm ~0.0138$ , $0.8937~\pm ~0.0121$ , and $0.9736~\pm ~0.0025$ , respectively, confirming its robustness and adaptability across diverse real-world HAR scenarios.
- Open Access
- 13 September 2023
- Published by MDPI AG in Bioengineering
- Vol. 10 (9) , 1082
Multivariate CNN Model for Human Locomotion Activity Recognition with a Wearable Exoskeleton Robot
This study introduces a novel convolutional neural network (CNN) architecture, encompassing both single and multi-head designs, developed to identify a user’s locomotion activity while using a wearable lower limb robot. Our research involved 500 healthy adult participants in an activities of daily living (ADL) space, conducted from 1 September to 30 November 2022. We collected prospective data to identify five locomotion activities (level ground walking, stair ascent/descent, and ramp ascent/descent) across three terrains: flat ground, staircase, and ramp. To evaluate the predictive capabilities of the proposed CNN architectures, we compared its performance with three other models: one CNN and two hybrid models (CNN-LSTM and LSTM-CNN). Experiments were conducted using multivariate signals of various types obtained from electromyograms (EMGs) and the wearable robot. Our results reveal that the deeper CNN architecture significantly surpasses the performance of the three competing models. The proposed model, leveraging encoder data such as hip angles and velocities, along with postural signals such as roll, pitch, and yaw from the wearable lower limb robot, achieved superior performance with an inference speed of 1.14 s. Specifically, the F-measure performance of the proposed model reached 96.17%, compared to 90.68% for DDLMI, 94.41% for DeepConvLSTM, and 95.57% for LSTM-CNN, respectively.
- 13 September 2023
- Published by MDPI AG in Bioengineering
- Vol. 10 (9) , 1082
Open AccessMultivariate CNN Model for Human Locomotion Activity Recognition with a Wearable Exoskeleton Robot
This study introduces a novel convolutional neural network (CNN) architecture, encompassing both single and multi-head designs, developed to identify a user’s locomotion activity while using a wearable lower limb robot. Our research involved 500 healthy adult participants in an activities of daily living (ADL) space, conducted from 1 September to 30 November 2022. We collected prospective data to identify five locomotion activities (level ground walking, stair ascent/descent, and ramp ascent/descent) across three terrains: flat ground, staircase, and ramp. To evaluate the predictive capabilities of the proposed CNN architectures, we compared its performance with three other models: one CNN and two hybrid models (CNN-LSTM and LSTM-CNN). Experiments were conducted using multivariate signals of various types obtained from electromyograms (EMGs) and the wearable robot. Our results reveal that the deeper CNN architecture significantly surpasses the performance of the three competing models. The proposed model, leveraging encoder data such as hip angles and velocities, along with postural signals such as roll, pitch, and yaw from the wearable lower limb robot, achieved superior performance with an inference speed of 1.14 s. Specifically, the F-measure performance of the proposed model reached 96.17%, compared to 90.68% for DDLMI, 94.41% for DeepConvLSTM, and 95.57% for LSTM-CNN, respectively.
- Open Access
Propensity-Score-Matched Evaluation of Adverse Events Affecting Recovery after COVID-19 Vaccination: On Adenovirus and mRNA Vaccines
This study aimed to observe adverse events following immunisation (AEFIs) that affected recovery within two weeks after COVID-19 vaccination and investigate their risks in propensity-score-matched populations. Data were collected from 447,346 reports from the VAERS between 1 January 2021 and 31 July 2021. Propensity-score-matched populations were constructed by adjusting for demographic characteristics and 11 underlying diseases in eligible subjects who received 1 of 3 COVID-19 vaccines: 19,462 Ad26.COV2.S, 120,580 mRNA-1273, and 100,752 BNT162b2. We observed that 88 suspected AEFIs (22 in Ad26.COV2.S, 62 in mRNA-1273, and 54 in BNT162b2) were associated with an increased risk of delayed recovery within 2 weeks after COVID-19 vaccinations. Nervous system, musculoskeletal and connective tissue, gastrointestinal, skin, and subcutaneous tissue disorders were the most common AEFIs after COVID-19 vaccination. Interestingly, four local and systemic reactions affected recovery in different vaccine recipients during our study period: asthenic conditions and febrile disorders in Ad26.COV2.S and mRNA-1273; general signs and symptoms in mRNA-1273 and BNT162b2; injection site reactions in Ad26.COV2.S and BNT162b2. Although it is necessary to confirm a causal relationship with COVID-19 vaccinations, some symptoms, including paralysis, allergic disorders, breathing abnormalities, and visual impairment, may hinder the recovery of these recipients.
Open AccessPropensity-Score-Matched Evaluation of Adverse Events Affecting Recovery after COVID-19 Vaccination: On Adenovirus and mRNA Vaccines
This study aimed to observe adverse events following immunisation (AEFIs) that affected recovery within two weeks after COVID-19 vaccination and investigate their risks in propensity-score-matched populations. Data were collected from 447,346 reports from the VAERS between 1 January 2021 and 31 July 2021. Propensity-score-matched populations were constructed by adjusting for demographic characteristics and 11 underlying diseases in eligible subjects who received 1 of 3 COVID-19 vaccines: 19,462 Ad26.COV2.S, 120,580 mRNA-1273, and 100,752 BNT162b2. We observed that 88 suspected AEFIs (22 in Ad26.COV2.S, 62 in mRNA-1273, and 54 in BNT162b2) were associated with an increased risk of delayed recovery within 2 weeks after COVID-19 vaccinations. Nervous system, musculoskeletal and connective tissue, gastrointestinal, skin, and subcutaneous tissue disorders were the most common AEFIs after COVID-19 vaccination. Interestingly, four local and systemic reactions affected recovery in different vaccine recipients during our study period: asthenic conditions and febrile disorders in Ad26.COV2.S and mRNA-1273; general signs and symptoms in mRNA-1273 and BNT162b2; injection site reactions in Ad26.COV2.S and BNT162b2. Although it is necessary to confirm a causal relationship with COVID-19 vaccinations, some symptoms, including paralysis, allergic disorders, breathing abnormalities, and visual impairment, may hinder the recovery of these recipients.
- Open Access
- 5 October 2021
- Published by Oxford University Press (OUP) in Cerebral Cortex Communications
- Vol. 2 (4) , tgab058
Time Course of Odor Categorization Processing
The brain’s mechanisms for categorizing different odors have long been a research focus. Previous studies suggest that odor categorization may involve multiple neurological processes within the brain with temporal and spatial neuronal activation. However, there is limited evidence regarding temporally mediated mechanisms in humans, especially millisecond odor processing. Such mechanisms may be important because different brain areas may play different roles at a particular activation time during sensory processing. Here, we focused on how the brain categorizes odors at specific time intervals. Using multivariate electroencephalography (EEG) analysis, we found that similarly perceived odors induced similar EEG signals during 50–100, 150–200, and 350–400 ms at the theta frequency. We also found significant activation at 100–150 and 350–400 ms at the gamma frequency. At these two frequencies, significant activation was observed in some olfactory-associated areas, including the orbitofrontal cortex. Our findings provide essential evidence that specific periods may be related to odor quality processing during central olfactory processing.
- 5 October 2021
- Published by Oxford University Press (OUP) in Cerebral Cortex Communications
- Vol. 2 (4) , tgab058
Open AccessTime Course of Odor Categorization Processing
The brain’s mechanisms for categorizing different odors have long been a research focus. Previous studies suggest that odor categorization may involve multiple neurological processes within the brain with temporal and spatial neuronal activation. However, there is limited evidence regarding temporally mediated mechanisms in humans, especially millisecond odor processing. Such mechanisms may be important because different brain areas may play different roles at a particular activation time during sensory processing. Here, we focused on how the brain categorizes odors at specific time intervals. Using multivariate electroencephalography (EEG) analysis, we found that similarly perceived odors induced similar EEG signals during 50–100, 150–200, and 350–400 ms at the theta frequency. We also found significant activation at 100–150 and 350–400 ms at the gamma frequency. At these two frequencies, significant activation was observed in some olfactory-associated areas, including the orbitofrontal cortex. Our findings provide essential evidence that specific periods may be related to odor quality processing during central olfactory processing.
- Open Access
- 1 October 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 26 (4) , 1802-1814
Machine Learning to Identify Psychomotor Behaviors of Delirium for Patients in Long-Term Care Facility
This study aimed to develop accurate and explainable machine learning models for three psychomotor behaviors of delirium for hospitalized adult patients. A prospective pilot study was conducted with 33 participants admitted to a long-term care facility between August 10 and 25, 2020. During the pilot study, we collected 560 cases that included 33 clinical variables and the survey items from the short confusion assessment method (S-CAM), and developed a mobile-based application. Multiple machine learning algorithms, including four rule-mining algorithms (C4.5, CBA, MCAR, and LEM2) and four other statistical learning algorithms (LR, ANNs, SVMs with three kernel functions, and random forest), were validated by paired Wilcoxon signed-rank tests on both macro-averaged F1 and weighted average F1-measures during the 10-times stratified 2-fold cross-validation. The LEM2 algorithm achieved the best prediction performance (macro-averaged F1-measure of 49.35%; weighted average F1-measure of 96.55%), correctly identifying adult patients at delirium risk. In the pairwise comparison between predictive powers observed from independent models, the LEM2 model showed a medium or large effect size between 0.4925 and 0.8766 when compared with LR, ANN, SVM with RBF, and MCAR models. We have confirmed that acute consciousness in S-CAM assessment is closely associated with different predictors for screening three psychomotor behaviors of delirium: 1) education level, dementia type or its level, sleep disorder, dehydration, and infection in mixed-type delirium; 2) gender, education level, dementia type, dehydration, bedsores, and foley catheter in hyperactive delirium; and 3) pain, sleep disorder, and haloperidol use in hypoactive delirium.
- 1 October 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 26 (4) , 1802-1814
Open AccessMachine Learning to Identify Psychomotor Behaviors of Delirium for Patients in Long-Term Care Facility
This study aimed to develop accurate and explainable machine learning models for three psychomotor behaviors of delirium for hospitalized adult patients. A prospective pilot study was conducted with 33 participants admitted to a long-term care facility between August 10 and 25, 2020. During the pilot study, we collected 560 cases that included 33 clinical variables and the survey items from the short confusion assessment method (S-CAM), and developed a mobile-based application. Multiple machine learning algorithms, including four rule-mining algorithms (C4.5, CBA, MCAR, and LEM2) and four other statistical learning algorithms (LR, ANNs, SVMs with three kernel functions, and random forest), were validated by paired Wilcoxon signed-rank tests on both macro-averaged F1 and weighted average F1-measures during the 10-times stratified 2-fold cross-validation. The LEM2 algorithm achieved the best prediction performance (macro-averaged F1-measure of 49.35%; weighted average F1-measure of 96.55%), correctly identifying adult patients at delirium risk. In the pairwise comparison between predictive powers observed from independent models, the LEM2 model showed a medium or large effect size between 0.4925 and 0.8766 when compared with LR, ANN, SVM with RBF, and MCAR models. We have confirmed that acute consciousness in S-CAM assessment is closely associated with different predictors for screening three psychomotor behaviors of delirium: 1) education level, dementia type or its level, sleep disorder, dehydration, and infection in mixed-type delirium; 2) gender, education level, dementia type, dehydration, bedsores, and foley catheter in hyperactive delirium; and 3) pain, sleep disorder, and haloperidol use in hypoactive delirium.
- Open Access
Joint Maximum Likelihood Time Delay Estimation of Unknown Event-Related Potential Signals for EEG Sensor Signal Quality Enhancement
Electroencephalograms (EEGs) measure a brain signal that contains abundant information about the human brain function and health. For this reason, recent clinical brain research and brain computer interface (BCI) studies use EEG signals in many applications. Due to the significant noise in EEG traces, signal processing to enhance the signal to noise power ratio (SNR) is necessary for EEG analysis, especially for non-invasive EEG. A typical method to improve the SNR is averaging many trials of event related potential (ERP) signal that represents a brain’s response to a particular stimulus or a task. The averaging, however, is very sensitive to variable delays. In this study, we propose two time delay estimation (TDE) schemes based on a joint maximum likelihood (ML) criterion to compensate the uncertain delays which may be different in each trial. We evaluate the performance for different types of signals such as random, deterministic, and real EEG signals. The results show that the proposed schemes provide better performance than other conventional schemes employing averaged signal as a reference, e.g., up to 4 dB gain at the expected delay error of 10°.
Open AccessJoint Maximum Likelihood Time Delay Estimation of Unknown Event-Related Potential Signals for EEG Sensor Signal Quality Enhancement
Electroencephalograms (EEGs) measure a brain signal that contains abundant information about the human brain function and health. For this reason, recent clinical brain research and brain computer interface (BCI) studies use EEG signals in many applications. Due to the significant noise in EEG traces, signal processing to enhance the signal to noise power ratio (SNR) is necessary for EEG analysis, especially for non-invasive EEG. A typical method to improve the SNR is averaging many trials of event related potential (ERP) signal that represents a brain’s response to a particular stimulus or a task. The averaging, however, is very sensitive to variable delays. In this study, we propose two time delay estimation (TDE) schemes based on a joint maximum likelihood (ML) criterion to compensate the uncertain delays which may be different in each trial. We evaluate the performance for different types of signals such as random, deterministic, and real EEG signals. The results show that the proposed schemes provide better performance than other conventional schemes employing averaged signal as a reference, e.g., up to 4 dB gain at the expected delay error of 10°.
- 28 July 2015
- Published by The Korea Society of Digital Policy and Management in Journal of Digital Convergence
- Vol. 13 (7) , 381-391
Impacts of Wellness Components on Individuals' Wellness Status for Wellness Convergence Systems
- 28 July 2015
- Published by The Korea Society of Digital Policy and Management in Journal of Digital Convergence
- Vol. 13 (7) , 381-391
Impacts of Wellness Components on Individuals' Wellness Status for Wellness Convergence Systems
- 30 June 2015
- Published by The Korea Society of Industrial Systems in Journal of the Korea Industrial Information Systems Research
- Vol. 20 (3) , 61-69
A probabilistic knowledge model for analyzing heart rate variability
- 30 June 2015
- Published by The Korea Society of Industrial Systems in Journal of the Korea Industrial Information Systems Research
- Vol. 20 (3) , 61-69
A probabilistic knowledge model for analyzing heart rate variability