Kyu-Hwan Jung
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
Kyu-Hwan Jung has a total of 158 co-authors. Here are the top co-authors by number of shared publications:
Jaemin Son
VUNO Inc., Seoul, Republic of Korea
9 shared publications- Sang Jun Park
Department of Ophthalmology, Seoul National University Bundang Hospital, Seoul National University College of Medicine, Seongnam, Korea
8 shared publications - Jaewook Lee
Seoul National University
6 shared publications - Joon Beom Seo
Department of Radiology and Research Institute of Radiology, University of Ulsan College of Medicine, Asan Medical Center, Seoul, Korea
6 shared publications Kyu Hyung Park
Department of Ophthalmology, Seoul National University College of Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea
5 shared publicationsJoo Young Shin
Department of Ophthalmology, Seoul Metropolitan Government Seoul National University Boramae Medical Centre, Seoul, Republic of Korea
4 shared publications
Latest Publications
- Open Access
- 11 August 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e92374
Finite State Machine–Guided Retrieval-Augmented Generation Improves Expert-Rated Acceptability of a Peripherally Inserted Central Catheter Self-Management Chatbot: Single-Center Content Validation Study
Background: Patients with cancer undergoing long-term or vesicant chemotherapy frequently require peripherally inserted central catheters (PICCs). Due to the nature of ambulatory treatment administration, self-PICC management is essential for the continuation and completion of the planned treatment. Large language models offer potential for continuous patient support, but hallucinations and insufficient adherence to clinical protocols remain concerns. Fine-tuning (FT) and retrieval-augmented generation (RAG) improve factual grounding but cannot enforce the structured decision logic of expert-led PICC consultations, leaving the value of dialogue-control mechanisms unclear. Objective: This study is an expert-panel evaluation designed to exploratorily verify whether a finite state machine (FSM)–guided RAG architecture yields incremental gains in clinical acceptability when added on top of FT and RAG compared with simpler architectures (fine-tuned model alone; fine-tuned model with RAG). Methods: We conducted a blinded ablation comparison of three chatbot architectures sharing an identical fine-tuned GPT-4o-mini base: model 1 (fine-tuning only; FT), model 2 (FT + RAG), and model 3 (FT + RAG + FSM). Three nurses specializing in PICC management (≥10 y experience) independently evaluated responses to 43 PICC-related queries. The two-stage evaluation comprised query-level forced-choice preference and global Likert ratings across 8 predefined dimensions. Interrater agreement was quantified with Gwet AC1. Preference rates were compared with Cochran Q and pairwise McNemar tests with Bonferroni correction. Global Likert ratings were summarized descriptively across 8 predefined dimensions. Results: The FSM-guided model (model 3) was preferred in 34 of 43 scenarios (79.1%). Interrater agreement was moderate (Gwet AC1=0.58, 95% CI 0.42‐0.75; P <.001). Differences in model preference were significant (Cochran Q=48.2; P <.001). Pairwise McNemar tests showed model 3 was preferred significantly more often than model 1 and model 2 (both P <.001), while model 1 versus model 2 did not differ after Bonferroni correction. For the global ratings, descriptive statistics indicated that model 3 received higher ratings than models 1 and 2 across most criteria, although it tended to receive lower ratings for efficiency, possibly reflecting its multiturn structure. In qualitative debriefing, the nurses noted occasional unnecessary conversational turns under FSM guidance. Conclusions: In this expert-panel content-validation study, the FSM-guided fine-tuned RAG model received higher expert-rated acceptability than the simpler architectures, with FSM-based dialogue control more closely aligning chatbot outputs with expert-led PICC consultation patterns. Patient-facing usability testing and multicenter validation are required before claims regarding patient safety or clinical deployment can be made.
- 11 August 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e92374
Open AccessFinite State Machine–Guided Retrieval-Augmented Generation Improves Expert-Rated Acceptability of a Peripherally Inserted Central Catheter Self-Management Chatbot: Single-Center Content Validation Study
Background: Patients with cancer undergoing long-term or vesicant chemotherapy frequently require peripherally inserted central catheters (PICCs). Due to the nature of ambulatory treatment administration, self-PICC management is essential for the continuation and completion of the planned treatment. Large language models offer potential for continuous patient support, but hallucinations and insufficient adherence to clinical protocols remain concerns. Fine-tuning (FT) and retrieval-augmented generation (RAG) improve factual grounding but cannot enforce the structured decision logic of expert-led PICC consultations, leaving the value of dialogue-control mechanisms unclear. Objective: This study is an expert-panel evaluation designed to exploratorily verify whether a finite state machine (FSM)–guided RAG architecture yields incremental gains in clinical acceptability when added on top of FT and RAG compared with simpler architectures (fine-tuned model alone; fine-tuned model with RAG). Methods: We conducted a blinded ablation comparison of three chatbot architectures sharing an identical fine-tuned GPT-4o-mini base: model 1 (fine-tuning only; FT), model 2 (FT + RAG), and model 3 (FT + RAG + FSM). Three nurses specializing in PICC management (≥10 y experience) independently evaluated responses to 43 PICC-related queries. The two-stage evaluation comprised query-level forced-choice preference and global Likert ratings across 8 predefined dimensions. Interrater agreement was quantified with Gwet AC1. Preference rates were compared with Cochran Q and pairwise McNemar tests with Bonferroni correction. Global Likert ratings were summarized descriptively across 8 predefined dimensions. Results: The FSM-guided model (model 3) was preferred in 34 of 43 scenarios (79.1%). Interrater agreement was moderate (Gwet AC1=0.58, 95% CI 0.42‐0.75; P <.001). Differences in model preference were significant (Cochran Q=48.2; P <.001). Pairwise McNemar tests showed model 3 was preferred significantly more often than model 1 and model 2 (both P <.001), while model 1 versus model 2 did not differ after Bonferroni correction. For the global ratings, descriptive statistics indicated that model 3 received higher ratings than models 1 and 2 across most criteria, although it tended to receive lower ratings for efficiency, possibly reflecting its multiturn structure. In qualitative debriefing, the nurses noted occasional unnecessary conversational turns under FSM guidance. Conclusions: In this expert-panel content-validation study, the FSM-guided fine-tuned RAG model received higher expert-rated acceptability than the simpler architectures, with FSM-based dialogue control more closely aligning chatbot outputs with expert-led PICC consultation patterns. Patient-facing usability testing and multicenter validation are required before claims regarding patient safety or clinical deployment can be made.
- Open Access
A generalizable foundation model for intraoperative understanding across surgical procedures
In minimally invasive surgery, clinical decisions depend on real-time visual interpretation, yet intraoperative perception varies substantially across surgeons and procedures. This variability limits consistent assessment, training, and the development of reliable artificial intelligence systems, as most surgical AI models are designed for narrowly defined tasks and do not generalize across procedures or institutions. Here we introduce ZEN, a generalizable foundation model for intraoperative surgical video understanding trained on more than 4 million frames from over 21 procedures using a self-supervised multi-teacher distillation framework. We curated a large and diverse dataset and systematically evaluated multiple representation learning strategies within a unified benchmark. Across 20 downstream tasks and full fine-tuning, frozen-backbone, few-shot and zero-shot settings, ZEN consistently outperforms existing surgical foundation models and demonstrates robust cross-procedure generalization. These results suggest a step toward unified representations for surgical scene understanding and support future applications in intraoperative assistance and surgical training assessment.
Open AccessA generalizable foundation model for intraoperative understanding across surgical procedures
In minimally invasive surgery, clinical decisions depend on real-time visual interpretation, yet intraoperative perception varies substantially across surgeons and procedures. This variability limits consistent assessment, training, and the development of reliable artificial intelligence systems, as most surgical AI models are designed for narrowly defined tasks and do not generalize across procedures or institutions. Here we introduce ZEN, a generalizable foundation model for intraoperative surgical video understanding trained on more than 4 million frames from over 21 procedures using a self-supervised multi-teacher distillation framework. We curated a large and diverse dataset and systematically evaluated multiple representation learning strategies within a unified benchmark. Across 20 downstream tasks and full fine-tuning, frozen-backbone, few-shot and zero-shot settings, ZEN consistently outperforms existing surgical foundation models and demonstrates robust cross-procedure generalization. These results suggest a step toward unified representations for surgical scene understanding and support future applications in intraoperative assistance and surgical training assessment.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- p. 1
CurConMix+: A Unified Spatio-Temporal Framework for Hierarchical Surgical Workflow Understanding
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- p. 1
Open AccessCurConMix+: A Unified Spatio-Temporal Framework for Hierarchical Surgical Workflow Understanding
- Open Access
- 1 January 2026
- Published by XMLink in Annals of Surgical Treatment and Research
- Vol. 110 (3) , 135-143
Physical AI goes to the operating room: are we ready for the Surgical Data Factory?
Oh N, et al. Ann Surg Treat Res. 2026 Mar;110(3):135-143. https://doi.org/10.4174/astr.2026.110.3.135
- 1 January 2026
- Published by XMLink in Annals of Surgical Treatment and Research
- Vol. 110 (3) , 135-143
Open AccessPhysical AI goes to the operating room: are we ready for the Surgical Data Factory?
Oh N, et al. Ann Surg Treat Res. 2026 Mar;110(3):135-143. https://doi.org/10.4174/astr.2026.110.3.135
- 21 September 2025
- Published by Springer Nature
- p. 149-158
CurConMix: A Curriculum Contrastive Learning Framework for Enhancing Surgical Action Triplet Recognition
- 21 September 2025
- Published by Springer Nature
- p. 149-158
CurConMix: A Curriculum Contrastive Learning Framework for Enhancing Surgical Action Triplet Recognition
- Top Cited PapersOpen Access
- 30 April 2025
- Published by The Korean Society of Medical Informatics in Healthcare Informatics Research
- Vol. 31 (2) , 114-124
Large Language Models in Medicine: Clinical Applications, Technical Challenges, and Ethical Considerations
Objectives: This study presents a comprehensive review of the clinical applications, technical challenges, and ethical considerations associated with using large language models (LLMs) in medicine.Methods: A literature survey of peer-reviewed articles, technical reports, and expert commentary from relevant medical and artificial intelligence journals was conducted. Key clinical application areas, technical limitations (e.g., accuracy, validation, transparency), and ethical issues (e.g., bias, safety, accountability, privacy) were identified and analyzed.Results: LLMs have potential in clinical documentation assistance, decision support, patient communication, and workflow optimization. The level of supporting evidence varies; documentation support applications are relatively mature, whereas autonomous diagnostics continue to face notable limitations regarding accuracy and validation. Key technical challenges include model hallucination, lack of robust clinical validation, integration issues, and limited transparency. Ethical concerns involve algorithmic bias risking health inequities, threats to patient safety from inaccuracies, unclear accountability, data privacy, and impacts on clinician-patient interactions.Conclusions: LLMs possess transformative potential for clinical medicine, particularly by augmenting clinician capabilities. However, substantial technical and ethical hurdles necessitate rigorous research, validation, clearly defined guidelines, and human oversight. Existing evidence supports an assistive rather than autonomous role, mandating careful, evidence-based integration that prioritizes patient safety and equity.
- 30 April 2025
- Published by The Korean Society of Medical Informatics in Healthcare Informatics Research
- Vol. 31 (2) , 114-124
Top Cited PapersOpen AccessLarge Language Models in Medicine: Clinical Applications, Technical Challenges, and Ethical Considerations
Objectives: This study presents a comprehensive review of the clinical applications, technical challenges, and ethical considerations associated with using large language models (LLMs) in medicine.Methods: A literature survey of peer-reviewed articles, technical reports, and expert commentary from relevant medical and artificial intelligence journals was conducted. Key clinical application areas, technical limitations (e.g., accuracy, validation, transparency), and ethical issues (e.g., bias, safety, accountability, privacy) were identified and analyzed.Results: LLMs have potential in clinical documentation assistance, decision support, patient communication, and workflow optimization. The level of supporting evidence varies; documentation support applications are relatively mature, whereas autonomous diagnostics continue to face notable limitations regarding accuracy and validation. Key technical challenges include model hallucination, lack of robust clinical validation, integration issues, and limited transparency. Ethical concerns involve algorithmic bias risking health inequities, threats to patient safety from inaccuracies, unclear accountability, data privacy, and impacts on clinician-patient interactions.Conclusions: LLMs possess transformative potential for clinical medicine, particularly by augmenting clinician capabilities. However, substantial technical and ethical hurdles necessitate rigorous research, validation, clearly defined guidelines, and human oversight. Existing evidence supports an assistive rather than autonomous role, mandating careful, evidence-based integration that prioritizes patient safety and equity.
- Open Access
- 11 December 2024
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 26 (1) , e63892
Leveraging Large Language Models for Improved Understanding of Communications With Patients With Cancer in a Call Center Setting: Proof-of-Concept Study
Background: Hospital call centers play a critical role in providing support and information to patients with cancer, making it crucial to effectively identify and understand patient intent during consultations. However, operational efficiency and standardization of telephone consultations, particularly when categorizing diverse patient inquiries, remain significant challenges. While traditional deep learning models like long short-term memory (LSTM) and bidirectional encoder representations from transformers (BERT) have been used to address these issues, they heavily depend on annotated datasets, which are labor-intensive and time-consuming to generate. Large language models (LLMs) like GPT-4, with their in-context learning capabilities, offer a promising alternative for classifying patient intent without requiring extensive retraining. Objective: This study evaluates the performance of GPT-4 in classifying the purpose of telephone consultations of patients with cancer. In addition, it compares the performance of GPT-4 to that of discriminative models, such as LSTM and BERT, with a particular focus on their ability to manage ambiguous and complex queries. Methods: We used a dataset of 430,355 sentences from telephone consultations with patients with cancer between 2016 and 2020. LSTM and BERT models were trained on 300,000 sentences using supervised learning, while GPT-4 was applied using zero-shot and few-shot approaches without explicit retraining. The accuracy of each model was compared using 1,000 randomly selected sentences from 2020 onward, with special attention paid to how each model handled ambiguous or uncertain queries. Results: GPT-4, which uses only a few examples (a few shots), attained a remarkable accuracy of 85.2%, considerably outperforming the LSTM and BERT models, which achieved accuracies of 73.7% and 71.3%, respectively. Notably, categories such as “Treatment,” “Rescheduling,” and “Symptoms” involve multiple contexts and exhibit significant complexity. GPT-4 demonstrated more than 15% superior performance in handling ambiguous queries in these categories. In addition, GPT-4 excelled in categories like “Records” and “Routine,” where contextual clues were clear, outperforming the discriminative models. These findings emphasize the potential of LLMs, particularly GPT-4, for interpreting complicated patient interactions during cancer-related telephone consultations. Conclusions: This study shows the potential of GPT-4 to significantly improve the classification of patient intent in cancer-related telephone oncological consultations. GPT-4’s ability to handle complex and ambiguous queries without extensive retraining provides a substantial advantage over discriminative models like LSTM and BERT. While GPT-4 demonstrates strong performance in various areas, further refinement of prompt design and category definitions is necessary to fully leverage its capabilities in practical health care applications. Future research will explore the integration of LLMs like GPT-4 into hybrid systems that combine human oversight with artificial intelligence–driven technologies.
- 11 December 2024
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 26 (1) , e63892
Open AccessLeveraging Large Language Models for Improved Understanding of Communications With Patients With Cancer in a Call Center Setting: Proof-of-Concept Study
Background: Hospital call centers play a critical role in providing support and information to patients with cancer, making it crucial to effectively identify and understand patient intent during consultations. However, operational efficiency and standardization of telephone consultations, particularly when categorizing diverse patient inquiries, remain significant challenges. While traditional deep learning models like long short-term memory (LSTM) and bidirectional encoder representations from transformers (BERT) have been used to address these issues, they heavily depend on annotated datasets, which are labor-intensive and time-consuming to generate. Large language models (LLMs) like GPT-4, with their in-context learning capabilities, offer a promising alternative for classifying patient intent without requiring extensive retraining. Objective: This study evaluates the performance of GPT-4 in classifying the purpose of telephone consultations of patients with cancer. In addition, it compares the performance of GPT-4 to that of discriminative models, such as LSTM and BERT, with a particular focus on their ability to manage ambiguous and complex queries. Methods: We used a dataset of 430,355 sentences from telephone consultations with patients with cancer between 2016 and 2020. LSTM and BERT models were trained on 300,000 sentences using supervised learning, while GPT-4 was applied using zero-shot and few-shot approaches without explicit retraining. The accuracy of each model was compared using 1,000 randomly selected sentences from 2020 onward, with special attention paid to how each model handled ambiguous or uncertain queries. Results: GPT-4, which uses only a few examples (a few shots), attained a remarkable accuracy of 85.2%, considerably outperforming the LSTM and BERT models, which achieved accuracies of 73.7% and 71.3%, respectively. Notably, categories such as “Treatment,” “Rescheduling,” and “Symptoms” involve multiple contexts and exhibit significant complexity. GPT-4 demonstrated more than 15% superior performance in handling ambiguous queries in these categories. In addition, GPT-4 excelled in categories like “Records” and “Routine,” where contextual clues were clear, outperforming the discriminative models. These findings emphasize the potential of LLMs, particularly GPT-4, for interpreting complicated patient interactions during cancer-related telephone consultations. Conclusions: This study shows the potential of GPT-4 to significantly improve the classification of patient intent in cancer-related telephone oncological consultations. GPT-4’s ability to handle complex and ambiguous queries without extensive retraining provides a substantial advantage over discriminative models like LSTM and BERT. While GPT-4 demonstrates strong performance in various areas, further refinement of prompt design and category definitions is necessary to fully leverage its capabilities in practical health care applications. Future research will explore the integration of LLMs like GPT-4 into hybrid systems that combine human oversight with artificial intelligence–driven technologies.
- Open Access
- 9 October 2024
- Published by Springer Nature in Scientific Data
- Vol. 11 (1) , 1-6
A large multi-focus dataset for white blood cell classification
The White Blood Cell (WBC) differential test ranks as the second most frequently performed diagnostic assay. It requires manual confirmation of the peripheral blood smear by experts to identify signs of abnormalities. Automated digital microscopy has emerged as a solution to reduce this labor-intensive process and improve efficiency. Several publicly available datasets provide various WBC subtypes of differing quality and resolution. These datasets have contributed to advancing WBC classification using machine learning techniques. However, digital microscopy of blood cells with high magnification often requires a wider depth of field, posing challenges for automatic digital microscopy that necessitates capturing multiple stacks of focal planes to obtain complete images of specific blood cells. Our dataset provides 25,773 image stacks from 72 patients. The image labels consist of 18 classes encompassing normal and abnormal cells, with two experts reviewing each label. Each image includes 10 z-stacks of cropped 200 by 200 pixel images, captured using a 50X microscope with 400 nm intervals. This study presents a comprehensive multi-focus dataset for WBC classification.
- 9 October 2024
- Published by Springer Nature in Scientific Data
- Vol. 11 (1) , 1-6
Open AccessA large multi-focus dataset for white blood cell classification
The White Blood Cell (WBC) differential test ranks as the second most frequently performed diagnostic assay. It requires manual confirmation of the peripheral blood smear by experts to identify signs of abnormalities. Automated digital microscopy has emerged as a solution to reduce this labor-intensive process and improve efficiency. Several publicly available datasets provide various WBC subtypes of differing quality and resolution. These datasets have contributed to advancing WBC classification using machine learning techniques. However, digital microscopy of blood cells with high magnification often requires a wider depth of field, posing challenges for automatic digital microscopy that necessitates capturing multiple stacks of focal planes to obtain complete images of specific blood cells. Our dataset provides 25,773 image stacks from 72 patients. The image labels consist of 18 classes encompassing normal and abnormal cells, with two experts reviewing each label. Each image includes 10 z-stacks of cropped 200 by 200 pixel images, captured using a 50X microscope with 400 nm intervals. This study presents a comprehensive multi-focus dataset for WBC classification.
- Open Access
- 1 October 2024
- Published by Association for Research in Vision and Ophthalmology (ARVO) in Translational Vision Science & Technology
- Vol. 13 (10) , 34
Clinical Utility of Deep Learning Assistance for Detecting Various Abnormal Findings in Color Retinal Fundus Images: A Reader Study
Purpose: To evaluate the clinical usefulness of a deep learning–based detection device for multiple abnormal findings on retinal fundus photographs for readers with varying expertise. Methods: Fourteen ophthalmologists (six residents, eight specialists) assessed 399 fundus images with respect to 12 major ophthalmologic findings, with or without the assistance of a deep learning algorithm, in two separate reading sessions. Sensitivity, specificity, and reading time per image were compared. Results: With algorithmic assistance, readers significantly improved in sensitivity for all 12 findings (P < 0.05) but tended to be less specific (P < 0.05) for hemorrhage, drusen, membrane, and vascular abnormality, more profoundly so in residents. Sensitivity without algorithmic assistance was significantly lower in residents (23.1%∼75.8%) compared to specialists (55.1%∼97.1%) in nine findings, but it improved to similar levels with algorithmic assistance (67.8%∼99.4% in residents, 83.2%∼99.5% in specialists) with only hemorrhage remaining statistically significantly lower. Variances in sensitivity were significantly reduced for all findings. Reading time per image decreased in images with fewer than three findings per image, more profoundly in residents. When simulated based on images acquired from a health screening center, average reading time was estimated to be reduced by 25.9% (from 16.4 seconds to 12.1 seconds per image) for residents, and by 2.0% (from 9.6 seconds to 9.4 seconds) for specialists. Conclusions: Deep learning–based computer-assisted detection devices increase sensitivity, reduce inter-reader variance in sensitivity, and reduce reading time in less complicated images. Translational Relevance: This study evaluated the influence that algorithmic assistance in detecting abnormal findings on retinal fundus photographs has on clinicians, possibly predicting its influence on clinical application.
- 1 October 2024
- Published by Association for Research in Vision and Ophthalmology (ARVO) in Translational Vision Science & Technology
- Vol. 13 (10) , 34
Open AccessClinical Utility of Deep Learning Assistance for Detecting Various Abnormal Findings in Color Retinal Fundus Images: A Reader Study
Purpose: To evaluate the clinical usefulness of a deep learning–based detection device for multiple abnormal findings on retinal fundus photographs for readers with varying expertise. Methods: Fourteen ophthalmologists (six residents, eight specialists) assessed 399 fundus images with respect to 12 major ophthalmologic findings, with or without the assistance of a deep learning algorithm, in two separate reading sessions. Sensitivity, specificity, and reading time per image were compared. Results: With algorithmic assistance, readers significantly improved in sensitivity for all 12 findings (P < 0.05) but tended to be less specific (P < 0.05) for hemorrhage, drusen, membrane, and vascular abnormality, more profoundly so in residents. Sensitivity without algorithmic assistance was significantly lower in residents (23.1%∼75.8%) compared to specialists (55.1%∼97.1%) in nine findings, but it improved to similar levels with algorithmic assistance (67.8%∼99.4% in residents, 83.2%∼99.5% in specialists) with only hemorrhage remaining statistically significantly lower. Variances in sensitivity were significantly reduced for all findings. Reading time per image decreased in images with fewer than three findings per image, more profoundly in residents. When simulated based on images acquired from a health screening center, average reading time was estimated to be reduced by 25.9% (from 16.4 seconds to 12.1 seconds per image) for residents, and by 2.0% (from 9.6 seconds to 9.4 seconds) for specialists. Conclusions: Deep learning–based computer-assisted detection devices increase sensitivity, reduce inter-reader variance in sensitivity, and reduce reading time in less complicated images. Translational Relevance: This study evaluated the influence that algorithmic assistance in detecting abnormal findings on retinal fundus photographs has on clinicians, possibly predicting its influence on clinical application.
- Open Access
- 30 September 2024
- Published by Sungkyunkwan University School of Medicine in Precision and Future Medicine
- Vol. 8 (3) , 92-104
Computationally efficient and stable real-world synthetic emergency room electronic health record data generation: high similarity and privacy preserving diffusion model approach: A retrospective cohort study
Purpose: This study aimed to develop real-world synthetic electronic health record (EHR) for emergency departments using computationally efficient and stable diffusion probabilistic models. Methods: In this study, we compared the performance of diffusion models and state-ofthe-art generative adversarial networks (GANs) in terms of statistical similarity, privacy, medical usefulness, and the feasibility of using synthetic data for machine learning purposes.Results: Our results demonstrate that diffusion models are significantly more computationally efficient than GANs and perform comparably or slightly better in terms of similarity, privacy, and utility. We also found that the data quality of the diffusion model is statistically very similar for both categorical and continuous values and can address class imbalance precisely. Moreover, the usefulness of synthetic data is almost identical to that of real EHR data. Our privacy analysis showed that the synthetic data generated by the diffusion models were private.Conclusion: These findings have significant implications for improving the efficiency of emergency settings and enabling real-time emergency room data modeling. This demonstrates the potential of diffusion models for generating computationally efficient high-quality synthetic data. The study concluded that diffusion models can generate real-world synthetic EHRs that are computationally efficient, private, and high-quality, and can be used for machine learning purposes in emergency settings.
- 30 September 2024
- Published by Sungkyunkwan University School of Medicine in Precision and Future Medicine
- Vol. 8 (3) , 92-104
Open AccessComputationally efficient and stable real-world synthetic emergency room electronic health record data generation: high similarity and privacy preserving diffusion model approach: A retrospective cohort study
Purpose: This study aimed to develop real-world synthetic electronic health record (EHR) for emergency departments using computationally efficient and stable diffusion probabilistic models. Methods: In this study, we compared the performance of diffusion models and state-ofthe-art generative adversarial networks (GANs) in terms of statistical similarity, privacy, medical usefulness, and the feasibility of using synthetic data for machine learning purposes.Results: Our results demonstrate that diffusion models are significantly more computationally efficient than GANs and perform comparably or slightly better in terms of similarity, privacy, and utility. We also found that the data quality of the diffusion model is statistically very similar for both categorical and continuous values and can address class imbalance precisely. Moreover, the usefulness of synthetic data is almost identical to that of real EHR data. Our privacy analysis showed that the synthetic data generated by the diffusion models were private.Conclusion: These findings have significant implications for improving the efficiency of emergency settings and enabling real-time emergency room data modeling. This demonstrates the potential of diffusion models for generating computationally efficient high-quality synthetic data. The study concluded that diffusion models can generate real-world synthetic EHRs that are computationally efficient, private, and high-quality, and can be used for machine learning purposes in emergency settings.