Joo Heung Yoon
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
Joo Heung Yoon has a total of 128 co-authors. Here are the top co-authors by number of shared publications:
Gilles Clermont
Department of Critical Care Medicine, School of Medicine, University of Pittsburgh, Pittsburgh, PA
6 shared publicationsMichael R. Pinsky
Department of Critical Care Medicine, University of Pittsburgh, Pittsburgh, USA
5 shared publicationsArtur Dubrawski
Carnegie Mellon University
4 shared publicationsBredy Pierre-Louis
Department of Medicine, Cardiology Division, New York Medical College, Valhalla, NY 10595, USA
3 shared publicationsChul Ahn
Department of Clinical Sciences, University of Texas Southwestern Medical Center, Dallas, TX
3 shared publications- Ik-Kyung Jang
Cardiology Division, Massachusetts General Hospital—Harvard Medical School, Boston, MA 02114, USA
3 shared publications
Latest Publications
- Open Access
- 15 July 2026
- Published by Springer Nature in Intensive Care Medicine Experimental
- Vol. 14 (1)
Leashing heterogeneous sepsis with phenotype-informed decision support modeling.
- 15 July 2026
- Published by Springer Nature in Intensive Care Medicine Experimental
- Vol. 14 (1)
Open AccessLeashing heterogeneous sepsis with phenotype-informed decision support modeling.
- Open Access
- 9 June 2026
- Published by Springer Nature in BMC Anesthesiology
- Vol. 26 (1)
HeLP-BAG score: a novel data-driven scoring system for predicting post-operative 30-day mortality using 24-hour preoperative data
Accurate preoperative risk stratification remains challenging, as existing scoring systems are often complex, invasive, or limited to specific patient populations. We aimed to develop a simple, interpretable, and broadly applicable risk score to screen for 30-day postoperative mortality using routinely available variables. We developed the HeLP-BAG score using three large surgical cohorts. To reduce demographic bias, cohorts were matched by age and sex, yielding 24,617 male and female patients per dataset. Clinical data collected within 24 h before surgery were used as inputs, and 30-day postoperative mortality was defined as the outcome. Candidate variables were selected based on clinical relevance, statistical separation, monotonicity, noninvasiveness, and general applicability. Change points were identified using a change-point detection algorithm and refined through clinical consensus. Model performance was evaluated in a combined validation cohort and compared with ASA-PS, seven established scoring systems, and five machine learning models using the area under the receiver operating characteristic curve (AUROC). ICU admission was additionally assessed using a post-hoc weighting strategy. Ten variables were included in the final score: ICU admission, age, mean arterial pressure, heart rate, hemoglobin, lactate, arterial pH, blood urea nitrogen, albumin, and glucose. The combined validation cohort consisted of 98,468 patients, with a 30-day mortality rate of 1.47%. The HeLP-BAG score achieved an AUROC of 0.873, showing comparable performance to ASA-PS (0.865; p = 0.120) and higher performance than SAPS3 (0.831). Performance was particularly robust in high-risk subgroups, including emergency surgery (AUROC 0.843; p < 0.001 vs. ASA-PS) and older patients aged ≥ 65 years (0.866; p = 0.004 vs. ASA-PS). Mortality increased monotonically with higher scores, and web-based calculator was developed to facilitate bedside implementation. The HeLP-BAG score is a simple and interpretable tool for screening 30-day postoperative mortality, demonstrating consistent performance across diverse surgical populations.
- 9 June 2026
- Published by Springer Nature in BMC Anesthesiology
- Vol. 26 (1)
Open AccessHeLP-BAG score: a novel data-driven scoring system for predicting post-operative 30-day mortality using 24-hour preoperative data
Accurate preoperative risk stratification remains challenging, as existing scoring systems are often complex, invasive, or limited to specific patient populations. We aimed to develop a simple, interpretable, and broadly applicable risk score to screen for 30-day postoperative mortality using routinely available variables. We developed the HeLP-BAG score using three large surgical cohorts. To reduce demographic bias, cohorts were matched by age and sex, yielding 24,617 male and female patients per dataset. Clinical data collected within 24 h before surgery were used as inputs, and 30-day postoperative mortality was defined as the outcome. Candidate variables were selected based on clinical relevance, statistical separation, monotonicity, noninvasiveness, and general applicability. Change points were identified using a change-point detection algorithm and refined through clinical consensus. Model performance was evaluated in a combined validation cohort and compared with ASA-PS, seven established scoring systems, and five machine learning models using the area under the receiver operating characteristic curve (AUROC). ICU admission was additionally assessed using a post-hoc weighting strategy. Ten variables were included in the final score: ICU admission, age, mean arterial pressure, heart rate, hemoglobin, lactate, arterial pH, blood urea nitrogen, albumin, and glucose. The combined validation cohort consisted of 98,468 patients, with a 30-day mortality rate of 1.47%. The HeLP-BAG score achieved an AUROC of 0.873, showing comparable performance to ASA-PS (0.865; p = 0.120) and higher performance than SAPS3 (0.831). Performance was particularly robust in high-risk subgroups, including emergency surgery (AUROC 0.843; p < 0.001 vs. ASA-PS) and older patients aged ≥ 65 years (0.866; p = 0.004 vs. ASA-PS). Mortality increased monotonically with higher scores, and web-based calculator was developed to facilitate bedside implementation. The HeLP-BAG score is a simple and interpretable tool for screening 30-day postoperative mortality, demonstrating consistent performance across diverse surgical populations.
- Open Access
PULSE-ICU: A Pretrained Unified Long-Sequence Encoder for Multi-task Prediction in Intensive Care Units
Intensive care unit (ICU) data are highly irregular, heterogeneous, and temporally fragmented, posing challenges for generalizable clinical prediction. We present PULSE-ICU, a self-supervised foundation model that learns event-level ICU representations from large-scale EHR sequences without resampling or manual feature engineering. A unified embedding module encodes event identity, continuous values, units, and temporal attributes, while a Longformer-based encoder enables efficient modeling of long trajectories. PULSE-ICU was fine-tuned across 18 prediction tasks, including mortality, intervention forecasting, and phenotype identification, achieving strong performance across task types. External validation on eICU, HiRID, and P12 showed substantial improvements with minimal fine-tuning, demonstrating robustness to domain shift and variable constraints. These findings suggest that foundation-style modeling can improve data efficiency and adaptability, providing a scalable framework for ICU decision support across diverse clinical environments.
Open AccessPULSE-ICU: A Pretrained Unified Long-Sequence Encoder for Multi-task Prediction in Intensive Care Units
Intensive care unit (ICU) data are highly irregular, heterogeneous, and temporally fragmented, posing challenges for generalizable clinical prediction. We present PULSE-ICU, a self-supervised foundation model that learns event-level ICU representations from large-scale EHR sequences without resampling or manual feature engineering. A unified embedding module encodes event identity, continuous values, units, and temporal attributes, while a Longformer-based encoder enables efficient modeling of long trajectories. PULSE-ICU was fine-tuned across 18 prediction tasks, including mortality, intervention forecasting, and phenotype identification, achieving strong performance across task types. External validation on eICU, HiRID, and P12 showed substantial improvements with minimal fine-tuning, demonstrating robustness to domain shift and variable constraints. These findings suggest that foundation-style modeling can improve data efficiency and adaptability, providing a scalable framework for ICU decision support across diverse clinical environments.
- Open Access
- 31 October 2025
- Published by The Korean Society of Medical Informatics in Healthcare Informatics Research
- Vol. 31 (4) , 416-425
Utility of Multimodal Large Language Models in Analyzing Chest X-Rays with Incomplete Contextual Information
Objectives: Large language models (LLMs) are increasingly used in clinical practice, but their performance can deteriorate when radiology reports are incomplete. We evaluated whether multimodal LLMs (integrating text and images) could enhance accuracy and interpretability in chest radiography reports, thereby improving their utility for clinical decision support. Specifically, we aimed to assess the robustness of LLMs in generating accurate impressions from chest radiography reports when provided with incomplete data, and whether multimodal input could mitigate performance loss.Methods: We analyzed 300 radiology image–report pairs from the MIMIC-CXR database. Three LLMs—OpenFlamingo, MedFlamingo, IDEFICS—were tested in text-only and multimodal formats. Chest X-ray impressions were generated from complete text reports and then regenerated after systematically removing 20%, 50%, and 80% of the text. The effect of adding images was evaluated using chest X-rays, and model performance was compared using three statistical methods. Hallucination rates were quantified.Results: In the text-only setting, OpenFlamingo, MedFlamingo, and IDEFICS demonstrated comparable performance (ROUGE-L: 0.23 vs. 0.21 vs. 0.21; F1RadGraph: 0.20 vs. 0.16 vs. 0.16; F1CheXbert: 0.49 vs. 0.41 vs. 0.41), with OpenFlamingo performing best on complete text (p < 0.001). All models exhibited performance decline with incomplete data. However, multimodal input significantly improved the performance of MedFlamingo and IDEFICS (p < 0.001), equaling or surpassing OpenFlamingo even under incomplete text conditions. Regarding hallucination, MedFlamingo showed a lower false-negative rate in multimodal compared with unimodal use, while false-positive rates were similar.Conclusions: LLMs may produce suboptimal outputs when radiology data are incomplete, but multimodal LLMs enhance reliability and may strengthen clinical decision-making support.
- 31 October 2025
- Published by The Korean Society of Medical Informatics in Healthcare Informatics Research
- Vol. 31 (4) , 416-425
Open AccessUtility of Multimodal Large Language Models in Analyzing Chest X-Rays with Incomplete Contextual Information
Objectives: Large language models (LLMs) are increasingly used in clinical practice, but their performance can deteriorate when radiology reports are incomplete. We evaluated whether multimodal LLMs (integrating text and images) could enhance accuracy and interpretability in chest radiography reports, thereby improving their utility for clinical decision support. Specifically, we aimed to assess the robustness of LLMs in generating accurate impressions from chest radiography reports when provided with incomplete data, and whether multimodal input could mitigate performance loss.Methods: We analyzed 300 radiology image–report pairs from the MIMIC-CXR database. Three LLMs—OpenFlamingo, MedFlamingo, IDEFICS—were tested in text-only and multimodal formats. Chest X-ray impressions were generated from complete text reports and then regenerated after systematically removing 20%, 50%, and 80% of the text. The effect of adding images was evaluated using chest X-rays, and model performance was compared using three statistical methods. Hallucination rates were quantified.Results: In the text-only setting, OpenFlamingo, MedFlamingo, and IDEFICS demonstrated comparable performance (ROUGE-L: 0.23 vs. 0.21 vs. 0.21; F1RadGraph: 0.20 vs. 0.16 vs. 0.16; F1CheXbert: 0.49 vs. 0.41 vs. 0.41), with OpenFlamingo performing best on complete text (p < 0.001). All models exhibited performance decline with incomplete data. However, multimodal input significantly improved the performance of MedFlamingo and IDEFICS (p < 0.001), equaling or surpassing OpenFlamingo even under incomplete text conditions. Regarding hallucination, MedFlamingo showed a lower false-negative rate in multimodal compared with unimodal use, while false-positive rates were similar.Conclusions: LLMs may produce suboptimal outputs when radiology data are incomplete, but multimodal LLMs enhance reliability and may strengthen clinical decision-making support.
- Open Access
Mind the Missing: Variable-Aware Representation Learning for Irregular EHR Time Series using Large Language Models
Irregular sampling and high missingness are intrinsic challenges in modeling time series derived from electronic health records (EHRs),where clinical variables are measured at uneven intervals depending on workflow and intervention timing. To address this, we propose VITAL, a variable-aware, large language model (LLM) based framework tailored for learning from irregularly sampled physiological time series. VITAL differentiates between two distinct types of clinical variables: vital signs, which are frequently recorded and exhibit temporal patterns, and laboratory tests, which are measured sporadically and lack temporal structure. It reprograms vital signs into the language space, enabling the LLM to capture temporal context and reason over missing values through explicit encoding. In contrast, laboratory variables are embedded either using representative summary values or a learnable [Not measured] token, depending on their availability. Extensive evaluations on the benchmark datasets from the PhysioNet demonstrate that VITAL outperforms state of the art methods designed for irregular time series. Furthermore, it maintains robust performance under high levels of missingness, which is prevalent in real world clinical scenarios where key variables are often unavailable.
Open AccessMind the Missing: Variable-Aware Representation Learning for Irregular EHR Time Series using Large Language Models
Irregular sampling and high missingness are intrinsic challenges in modeling time series derived from electronic health records (EHRs),where clinical variables are measured at uneven intervals depending on workflow and intervention timing. To address this, we propose VITAL, a variable-aware, large language model (LLM) based framework tailored for learning from irregularly sampled physiological time series. VITAL differentiates between two distinct types of clinical variables: vital signs, which are frequently recorded and exhibit temporal patterns, and laboratory tests, which are measured sporadically and lack temporal structure. It reprograms vital signs into the language space, enabling the LLM to capture temporal context and reason over missing values through explicit encoding. In contrast, laboratory variables are embedded either using representative summary values or a learnable [Not measured] token, depending on their availability. Extensive evaluations on the benchmark datasets from the PhysioNet demonstrate that VITAL outperforms state of the art methods designed for irregular time series. Furthermore, it maintains robust performance under high levels of missingness, which is prevalent in real world clinical scenarios where key variables are often unavailable.
- 4 July 2025
- Published by Elsevier in Anesthesiology Clinics
- Vol. 43 (3) , 507-525
Current Perspectives on the Artificial Intelligence in Critical Care Medicine
- 4 July 2025
- Published by Elsevier in Anesthesiology Clinics
- Vol. 43 (3) , 507-525
Current Perspectives on the Artificial Intelligence in Critical Care Medicine
- 13 June 2024
- Published by Elsevier in The Lancet
- Vol. 403 (10444) , 2589
The South Korean health-care system in crisis
- 13 June 2024
- Published by Elsevier in The Lancet
- Vol. 403 (10444) , 2589
The South Korean health-care system in crisis
- Open Access
- 26 March 2024
- Published by Wolters Kluwer Health in Critical Care Explorations
- Vol. 6 (4) , e1073
Early Physiologic Numerical and Waveform Characteristics of Simulated Hemorrhagic Events With Healthy Volunteers Donating Blood
OBJECTIVES: Early signs of bleeding are often masked by the physiologic compensatory responses delaying its identification. We sought to describe early physiologic signatures of bleeding during the blood donation process.SETTING: Waveform-level vital sign data including electrocardiography, photoplethysmography (PPG), continuous noninvasive arterial pressure, and respiratory waveforms were collected before, during, and after bleeding.SUBJECTS: Fifty-five healthy volunteers visited blood donation center to donate whole blood.INTERVENTION: After obtaining the informed consent, 3 minutes of resting time was given to each subject. Then 3 minutes of orthostasis was done, followed by another 3 minutes of resting before the blood donation. After the completion of donating blood, another 3 minutes of postbleeding resting time, followed by 3 minutes of orthostasis period again.MEASUREMENTS AND MAIN RESULTS: From 55 subjects, waveform signals as well as numerical vital signs (heart rate [HR], respiratory rate, blood pressure) and clinical characteristics were collected, and data from 51 subjects were analyzable. Any adverse events (AEs; dizziness, lightheadedness, nausea) were documented. Statistical and physiologic features including HR variability (HRV) metrics and other waveform morphologic parameters were modeled. Feature trends for all participants across the study protocol were analyzed. No significant changes in HR, blood pressure, or estimated cardiac output were seen during bleeding. Both orthostatic challenges and bleeding significantly decreased time domain and high-frequency domain HRV, and PPG amplitude, whereas increasing PPG amplitude variation. During bleeding, time-domain HRV feature trends were most sensitive to the first 100 mL of blood loss, and incremental changes of different HRV parameters (from 300 mL of blood loss), as well as a PPG morphologic feature (from 400 mL of blood loss), were shown with statistical significance. The AE group (n= 6) showed decreased sample entropy compared with the non-AE group during postbleed orthostatic challenge (p= 0.003). No significant other trend differences were observed during bleeding between AE and non-AE groups.CONCLUSIONS: Various HRV-related features were changed during rapid bleeding seen within the first minute. Subjects with AE during postbleeding orthostasis showed decreased sample entropy. These findings could be leveraged toward earlier identification of donors at risk for AE, and more broadly building a data-driven hemorrhage model for the early treatment of critical bleeding.
- 26 March 2024
- Published by Wolters Kluwer Health in Critical Care Explorations
- Vol. 6 (4) , e1073
Open AccessEarly Physiologic Numerical and Waveform Characteristics of Simulated Hemorrhagic Events With Healthy Volunteers Donating Blood
OBJECTIVES: Early signs of bleeding are often masked by the physiologic compensatory responses delaying its identification. We sought to describe early physiologic signatures of bleeding during the blood donation process.SETTING: Waveform-level vital sign data including electrocardiography, photoplethysmography (PPG), continuous noninvasive arterial pressure, and respiratory waveforms were collected before, during, and after bleeding.SUBJECTS: Fifty-five healthy volunteers visited blood donation center to donate whole blood.INTERVENTION: After obtaining the informed consent, 3 minutes of resting time was given to each subject. Then 3 minutes of orthostasis was done, followed by another 3 minutes of resting before the blood donation. After the completion of donating blood, another 3 minutes of postbleeding resting time, followed by 3 minutes of orthostasis period again.MEASUREMENTS AND MAIN RESULTS: From 55 subjects, waveform signals as well as numerical vital signs (heart rate [HR], respiratory rate, blood pressure) and clinical characteristics were collected, and data from 51 subjects were analyzable. Any adverse events (AEs; dizziness, lightheadedness, nausea) were documented. Statistical and physiologic features including HR variability (HRV) metrics and other waveform morphologic parameters were modeled. Feature trends for all participants across the study protocol were analyzed. No significant changes in HR, blood pressure, or estimated cardiac output were seen during bleeding. Both orthostatic challenges and bleeding significantly decreased time domain and high-frequency domain HRV, and PPG amplitude, whereas increasing PPG amplitude variation. During bleeding, time-domain HRV feature trends were most sensitive to the first 100 mL of blood loss, and incremental changes of different HRV parameters (from 300 mL of blood loss), as well as a PPG morphologic feature (from 400 mL of blood loss), were shown with statistical significance. The AE group (n= 6) showed decreased sample entropy compared with the non-AE group during postbleed orthostatic challenge (p= 0.003). No significant other trend differences were observed during bleeding between AE and non-AE groups.CONCLUSIONS: Various HRV-related features were changed during rapid bleeding seen within the first minute. Subjects with AE during postbleeding orthostasis showed decreased sample entropy. These findings could be leveraged toward earlier identification of donors at risk for AE, and more broadly building a data-driven hemorrhage model for the early treatment of critical bleeding.
- Open Access
- 15 May 2023
- Published by The Korean Society of Emergency Medicine in Clinical and Experimental Emergency Medicine
- Vol. 10 (2) , 132-137
Current challenges in adopting machine learning to critical care and emergency medicine
Over the past decades, the field of machine learning (ML) has made great strides in medicine. Despite the number of ML-inspired publications in the clinical arena, the results and implications are not readily accepted at the bedside. Although ML is very powerful in deciphering hidden patterns in complex critical care and emergency medicine data, various factors including data, feature generation, model design, performance assessment, and limited implementation could affect the utility of the research. In this short review, a series of current challenges of adopting ML models to clinical research will be discussed.
- 15 May 2023
- Published by The Korean Society of Emergency Medicine in Clinical and Experimental Emergency Medicine
- Vol. 10 (2) , 132-137
Open AccessCurrent challenges in adopting machine learning to critical care and emergency medicine
Over the past decades, the field of machine learning (ML) has made great strides in medicine. Despite the number of ML-inspired publications in the clinical arena, the results and implications are not readily accepted at the bedside. Although ML is very powerful in deciphering hidden patterns in complex critical care and emergency medicine data, various factors including data, feature generation, model design, performance assessment, and limited implementation could affect the utility of the research. In this short review, a series of current challenges of adopting ML models to clinical research will be discussed.
- 31 March 2022
- Published by Springer Nature
- p. 353-367
Artificial Intelligence in Critical Care Medicine
- 31 March 2022
- Published by Springer Nature
- p. 353-367
Artificial Intelligence in Critical Care Medicine