Satoshi Kodera
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Satoshi Kodera has a total of 168 co-authors. Here are the top co-authors by number of shared publications:
Hiroyuki Morita
Department of Cardiovascular Medicine, The University of Tokyo Hospital , 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8655
13 shared publicationsIssei Komuro
Department of Frontier Cardiovascular Science, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan
10 shared publicationsMasataka Sato
Department of Cardiovascular Medicine The University of Tokyo Hospital Tokyo Japan
9 shared publicationsHiroki Shinohara
Department of Cardiovascular Medicine, The University of Tokyo Hospital, Tokyo, Japan
8 shared publicationsJunji Kanda
Department of Cardiovascular Medicine, Asahi General Hospital, Chiba, Japan
8 shared publicationsHiroshi Akazawa
Department of Cardiovascular Medicine, The University of Tokyo Hospital, Tokyo, Japan
7 shared publications
Latest Publications
- Open Access
- 22 July 2026
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 7 (7)
Large language model–based simulated patient training for heart failure palliative care communication: a pilot study
Aims: Heart failure (HF) palliative care communication is essential but difficult to train at scale because conventional role-play programmes require facilitators and standardized patients. Large language models (LLMs) have emerged as potential tools for scalable communication training. This pilot study aimed to evaluate the feasibility of a web-based LLM-driven communication training application and to explore its early educational signal on physicians’ self-efficacy. Methods and results: This single-arm pilot study included physicians who completed one session using a Japanese-language web-based LLM application designed to simulate patients with advanced HF and provide automated framework-based feedback. The primary outcome was change in self-efficacy scores assessed by pre- and post-session questionnaires. Ten sessions were analysed. Physicians engaged in a mean of 7.6 ± 2.0 dialogue turns. Mean response time per model output and feedback generation were approximately 3 and 17 s, respectively. Significant improvements were observed in knowledge of palliative care communication (mean difference +1.7, adjusted P < 0.01) and confidence in HF palliative care communication (+1.2, adjusted P = 0.03). Other domains showed non-significant changes. Conclusion: This pilot study demonstrated the feasibility of a web-based LLM-simulated patient system and suggested an early educational signal in physicians’ self-efficacy for HF palliative care communication. Our scalable LLM-driven communication training may complement traditional educational approaches with further evaluation in larger controlled studies. Clinical trial registration: Trial registration number: UMIN000059988
- 22 July 2026
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 7 (7)
Open AccessLarge language model–based simulated patient training for heart failure palliative care communication: a pilot study
Aims: Heart failure (HF) palliative care communication is essential but difficult to train at scale because conventional role-play programmes require facilitators and standardized patients. Large language models (LLMs) have emerged as potential tools for scalable communication training. This pilot study aimed to evaluate the feasibility of a web-based LLM-driven communication training application and to explore its early educational signal on physicians’ self-efficacy. Methods and results: This single-arm pilot study included physicians who completed one session using a Japanese-language web-based LLM application designed to simulate patients with advanced HF and provide automated framework-based feedback. The primary outcome was change in self-efficacy scores assessed by pre- and post-session questionnaires. Ten sessions were analysed. Physicians engaged in a mean of 7.6 ± 2.0 dialogue turns. Mean response time per model output and feedback generation were approximately 3 and 17 s, respectively. Significant improvements were observed in knowledge of palliative care communication (mean difference +1.7, adjusted P < 0.01) and confidence in HF palliative care communication (+1.2, adjusted P = 0.03). Other domains showed non-significant changes. Conclusion: This pilot study demonstrated the feasibility of a web-based LLM-simulated patient system and suggested an early educational signal in physicians’ self-efficacy for HF palliative care communication. Our scalable LLM-driven communication training may complement traditional educational approaches with further evaluation in larger controlled studies. Clinical trial registration: Trial registration number: UMIN000059988
- Open Access
- 10 June 2026
- Published by Elsevier in Journal of Cardiology
- Vol. 88 (5) , 353-360
From DNA to drug discovery AI models for cardiovascular precision medicine
- 10 June 2026
- Published by Elsevier in Journal of Cardiology
- Vol. 88 (5) , 353-360
Open AccessFrom DNA to drug discovery AI models for cardiovascular precision medicine
- Open Access
- 5 May 2026
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 7 (5)
Deep learning model for genotype prediction from echocardiographic videos in non-ischaemic dilated cardiomyopathy
Aims: Non-ischaemic dilated cardiomyopathy (DCM) is frequently characterized by the presence of pathogenic germline variants, and genotype positivity predicts poor prognosis. Despite its importance, genetic testing remains underutilized in the current era. Therefore, we aimed to develop a deep learning model to predict genotype positivity using echocardiographic videos. Methods and results: We included patients who were diagnosed with DCM and had genetic testing at the University of Tokyo Hospital, Japan, consecutively from 2014 to 2022. The apical four-chamber views of echocardiographic videos were collected. First, we developed a deep learning model based on the EchoNet-Dynamic model, and the area under the curve (AUC) was computed. Second, we calculated the Madrid genotype score (clinical scoring system) for each case. Third, we developed a logistic regression model that combined the Madrid genotype score and the deep learning model. Finally, we compared the AUC of the combined model with that of the Madrid genotype score alone by DeLong’s test. Out of the 258 patients, 117 patients (45.3%) had genetic variants, and 141 (54.7%) did not. TTN (30.8%) was the most common genotype, followed by LMNA (18.8%). The deep learning model yielded an AUC of 0.64. The Madrid genotype score was well validated and achieved an AUC of 0.73. The combined model yielded an AUC of 0.76 with a significant improvement from the Madrid genotype score alone (P = 0.03). Conclusion: The deep learning model demonstrated modest discriminative ability to predict genotype positivity using echocardiographic videos. The accuracy of the clinical scoring system improved when combined with the deep learning model.
- 5 May 2026
- Published by Oxford University Press (OUP) in European Heart Journal - Digital Health
- Vol. 7 (5)
Open AccessDeep learning model for genotype prediction from echocardiographic videos in non-ischaemic dilated cardiomyopathy
Aims: Non-ischaemic dilated cardiomyopathy (DCM) is frequently characterized by the presence of pathogenic germline variants, and genotype positivity predicts poor prognosis. Despite its importance, genetic testing remains underutilized in the current era. Therefore, we aimed to develop a deep learning model to predict genotype positivity using echocardiographic videos. Methods and results: We included patients who were diagnosed with DCM and had genetic testing at the University of Tokyo Hospital, Japan, consecutively from 2014 to 2022. The apical four-chamber views of echocardiographic videos were collected. First, we developed a deep learning model based on the EchoNet-Dynamic model, and the area under the curve (AUC) was computed. Second, we calculated the Madrid genotype score (clinical scoring system) for each case. Third, we developed a logistic regression model that combined the Madrid genotype score and the deep learning model. Finally, we compared the AUC of the combined model with that of the Madrid genotype score alone by DeLong’s test. Out of the 258 patients, 117 patients (45.3%) had genetic variants, and 141 (54.7%) did not. TTN (30.8%) was the most common genotype, followed by LMNA (18.8%). The deep learning model yielded an AUC of 0.64. The Madrid genotype score was well validated and achieved an AUC of 0.73. The combined model yielded an AUC of 0.76 with a significant improvement from the Madrid genotype score alone (P = 0.03). Conclusion: The deep learning model demonstrated modest discriminative ability to predict genotype positivity using echocardiographic videos. The accuracy of the clinical scoring system improved when combined with the deep learning model.
- Open Access
- 6 March 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (3) , e0344084
PAH-former: Transfer learning for efficient discovery of pulmonary arterial hypertension-associated genes
Background and Aims: Pulmonary arterial hypertension (PAH) is a severe disease with limited effective therapies, making the discovery of new therapeutic targets crucial. While single-cell RNA sequencing (sc-RNA seq) offers a powerful tool for this purpose, its application is hampered by the scarcity of patient samples. This study addresses the problem of how to efficiently identify novel, functionally relevant disease-associated genes from limited publicly available data. Methods: We employed transfer learning by fine-tuning Geneformer, a deep learning model, with public sc-RNA seq data from patients with PAH to create a specialized model called PAH-former. This model was used to perform in silico perturbation analysis to identify and rank candidate genes predicted to influence the disease state. For validation, we performed RNA interference-mediated knockdown of top novel candidate genes in human pulmonary artery endothelial cells and measured the expression of SRY-Box Transcription Factor 18 ( SOX18 ), a signature gene of pulmonary arterial hypertension. Results: In silico perturbation analysis identified 134 candidate genes whose deletion was predicted to shift cells towards a disease phenotype. These included known disease-related genes as well as many novel ones. Subsequent in vitro validation demonstrated that knockdown of the candidate genes resulted in a significant increase in the expression of SOX18 . Conclusions: Our novel platform, PAH-former, provides a powerful and broadly applicable strategy for disease-related gene discovery. This approach enables the identification and validation of new candidate genes from limited data, promising to advance cell-specific mechanistic insights and accelerate therapeutic development for rare diseases like PAH. (248/300 words).
- 6 March 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (3) , e0344084
Open AccessPAH-former: Transfer learning for efficient discovery of pulmonary arterial hypertension-associated genes
Background and Aims: Pulmonary arterial hypertension (PAH) is a severe disease with limited effective therapies, making the discovery of new therapeutic targets crucial. While single-cell RNA sequencing (sc-RNA seq) offers a powerful tool for this purpose, its application is hampered by the scarcity of patient samples. This study addresses the problem of how to efficiently identify novel, functionally relevant disease-associated genes from limited publicly available data. Methods: We employed transfer learning by fine-tuning Geneformer, a deep learning model, with public sc-RNA seq data from patients with PAH to create a specialized model called PAH-former. This model was used to perform in silico perturbation analysis to identify and rank candidate genes predicted to influence the disease state. For validation, we performed RNA interference-mediated knockdown of top novel candidate genes in human pulmonary artery endothelial cells and measured the expression of SRY-Box Transcription Factor 18 ( SOX18 ), a signature gene of pulmonary arterial hypertension. Results: In silico perturbation analysis identified 134 candidate genes whose deletion was predicted to shift cells towards a disease phenotype. These included known disease-related genes as well as many novel ones. Subsequent in vitro validation demonstrated that knockdown of the candidate genes resulted in a significant increase in the expression of SOX18 . Conclusions: Our novel platform, PAH-former, provides a powerful and broadly applicable strategy for disease-related gene discovery. This approach enables the identification and validation of new candidate genes from limited data, promising to advance cell-specific mechanistic insights and accelerate therapeutic development for rare diseases like PAH. (248/300 words).
- Open Access
Contrastive Learning for Multi Label ECG Classification with Jaccard Score Based Sigmoid Loss
Recent advances in large language models (LLMs) have enabled the development of multimodal medical AI. While models such as MedGemini achieve high accuracy on VQA tasks like USMLE MM, their performance on ECG based tasks remains limited, and some models, such as MedGemma, do not support ECG data at all. Interpreting ECGs is inherently challenging, and diagnostic accuracy can vary depending on the interpreter's experience. Although echocardiography provides rich diagnostic information, it requires specialized equipment and personnel, limiting its availability. In this study, we focus on constructing a robust ECG encoder for multimodal pretraining using real world hospital data. We employ SigLIP, a CLIP based model with a sigmoid based loss function enabling multi label prediction, and introduce a modified loss function tailored to the multi label nature of ECG data. Experiments demonstrate that incorporating medical knowledge in the language model and applying the modified loss significantly improve multi label ECG classification. To further enhance performance, we increase the embedding dimensionality and apply random cropping to mitigate data drift. Finally, per label analysis reveals which ECG findings are easier or harder to predict. Our study provides a foundational framework for developing medical models that utilize ECG data.
Open AccessContrastive Learning for Multi Label ECG Classification with Jaccard Score Based Sigmoid Loss
Recent advances in large language models (LLMs) have enabled the development of multimodal medical AI. While models such as MedGemini achieve high accuracy on VQA tasks like USMLE MM, their performance on ECG based tasks remains limited, and some models, such as MedGemma, do not support ECG data at all. Interpreting ECGs is inherently challenging, and diagnostic accuracy can vary depending on the interpreter's experience. Although echocardiography provides rich diagnostic information, it requires specialized equipment and personnel, limiting its availability. In this study, we focus on constructing a robust ECG encoder for multimodal pretraining using real world hospital data. We employ SigLIP, a CLIP based model with a sigmoid based loss function enabling multi label prediction, and introduce a modified loss function tailored to the multi label nature of ECG data. Experiments demonstrate that incorporating medical knowledge in the language model and applying the modified loss significantly improve multi label ECG classification. To further enhance performance, we increase the embedding dimensionality and apply random cropping to mitigate data drift. Finally, per label analysis reveals which ECG findings are easier or harder to predict. Our study provides a foundational framework for developing medical models that utilize ECG data.
- Open Access
- 31 January 2026
- Published by International Heart Journal (Japanese Heart Journal) in International Heart Journal
- Vol. 67 (1) , 61-64
Validity of Diagnoses of Aortic Dissection and Pulmonary Embolism in Hospitalized Patients Using the Japanese DPC Database
Acute aortic dissection (AD) and pulmonary thromboembolism (PE) are life-threatening conditions requiring timely diagnosis. In Japan, large-scale validation studies of administrative data for these diagnoses are limited. It is required to validate the accuracy of International Classification of Diseases, 10th Revision (ICD-10) codes for acute AD and PE in the Diagnosis Procedure Combination (DPC) database using medical chart review as the reference standard.We reviewed DPC data from the University of Tokyo Hospital (April 2018-March 2024). Patients with I710 (AD) or I260/I269 (PE) listed in key diagnosis fields were identified. Chart reviews confirmed diagnoses, and 199 randomly selected patients without AD/PE codes served as negative controls. Sensitivity, specificity, and positive predictive value (PPV) were calculated.Among 10,483 records, 54 had DPC-coded AD and 71 had PE. Chart review confirmed 47 cases of acute AD and 48 of acute PE. No false negatives were found among controls. For acute AD, sensitivity was 100.0% (95% CI: 92.5–100.0), specificity 96.6% (93.1–98.6), and PPV 87.0% (75.1–94.6). For acute PE, sensitivity was 100.0% (92.6–100.0), specificity 89.6% (84.9–93.3), and PPV 67.6% (55.5–78.2).ICD-10 codes in the DPC database accurately identify acute AD and PE, with particularly high sensitivity. While PPV for acute PE is lower, likely due to inclusion of chronic conditions, DPC data can be cautiously used for research and surveillance in academic hospital settings. Multicenter validation is needed to ensure broader generalizability.
- 31 January 2026
- Published by International Heart Journal (Japanese Heart Journal) in International Heart Journal
- Vol. 67 (1) , 61-64
Open AccessValidity of Diagnoses of Aortic Dissection and Pulmonary Embolism in Hospitalized Patients Using the Japanese DPC Database
Acute aortic dissection (AD) and pulmonary thromboembolism (PE) are life-threatening conditions requiring timely diagnosis. In Japan, large-scale validation studies of administrative data for these diagnoses are limited. It is required to validate the accuracy of International Classification of Diseases, 10th Revision (ICD-10) codes for acute AD and PE in the Diagnosis Procedure Combination (DPC) database using medical chart review as the reference standard.We reviewed DPC data from the University of Tokyo Hospital (April 2018-March 2024). Patients with I710 (AD) or I260/I269 (PE) listed in key diagnosis fields were identified. Chart reviews confirmed diagnoses, and 199 randomly selected patients without AD/PE codes served as negative controls. Sensitivity, specificity, and positive predictive value (PPV) were calculated.Among 10,483 records, 54 had DPC-coded AD and 71 had PE. Chart review confirmed 47 cases of acute AD and 48 of acute PE. No false negatives were found among controls. For acute AD, sensitivity was 100.0% (95% CI: 92.5–100.0), specificity 96.6% (93.1–98.6), and PPV 87.0% (75.1–94.6). For acute PE, sensitivity was 100.0% (92.6–100.0), specificity 89.6% (84.9–93.3), and PPV 67.6% (55.5–78.2).ICD-10 codes in the DPC database accurately identify acute AD and PE, with particularly high sensitivity. While PPV for acute PE is lower, likely due to inclusion of chronic conditions, DPC data can be cautiously used for research and surveillance in academic hospital settings. Multicenter validation is needed to ensure broader generalizability.
- Open Access
- 5 November 2025
- Published by Elsevier in International Journal of Cardiology Cardiovascular Risk and Prevention
- Vol. 27, 200540
Impact of osteoporosis risk on in-hospital mortality in older patients with acute myocardial infarction: Insight from Tokyo CCU network database
- 5 November 2025
- Published by Elsevier in International Journal of Cardiology Cardiovascular Risk and Prevention
- Vol. 27, 200540
Open AccessImpact of osteoporosis risk on in-hospital mortality in older patients with acute myocardial infarction: Insight from Tokyo CCU network database
- Open Access
- 25 September 2025
- Published by The Japanese Coronary Association in Journal of Coronary Artery Disease
- Vol. 31 (3) , 106-111
Emerging Role of AI in Cardiovascular Medicine
This review provides an up-to-date overview of recent advances in artificial intelligence (AI) applied to cardiovascular medicine, spanning technological innovations, clinical applications, and broader societal implications. Over the past several years, deep learning methods have been increasingly adopted across multiple diagnostic modalities, including electrocardiography, echocardiography, coronary computed tomography angiography (CCTA), and intravascular ultrasound (IVUS), resulting in measurable improvements in diagnostic accuracy and workflow efficiency. In CCTA, AI facilitates automated detection of coronary stenosis and plaque burden. Echocardiography enables image segmentation and quantification of the left ventricular ejection fraction, while in IVUS, it is primarily applied to vessel structure analysis and lumen segmentation, and has also been used in conjunction with high-sensitivity troponin I to develop models capable of early ischemia assessment, offering the potential to rule out myocardial infarction from a single blood draw and to provide long-term risk estimates. More recently, domain-specific large language models, such as Med-PaLM, Gemini, and Articulate Medical Intelligence Explorer, have demonstrated diagnostic capabilities approaching physician-level performance in selected tasks. These systems are beginning to show promise for real-world clinical use by integrating multimodal data and providing interactive decision-making support. Nonetheless, several unresolved challenges persist, including the differences in cognitive reasoning between AI and clinicians, limited explainability, and complex regulatory and ethical considerations. Ongoing efforts, including randomized controlled trials (e.g., PROTEUS and RAPIDxAI) as well as the development of inclusive and sustainable digital health frameworks, aim to support the safe integration of AI into routine cardiovascular care. This review synthesizes current progress from technological, clinical, and implementation standpoints and explores the conditions under which AI can serve not as a replacement for clinicians, but as a collaborative partner in patient care.
- 25 September 2025
- Published by The Japanese Coronary Association in Journal of Coronary Artery Disease
- Vol. 31 (3) , 106-111
Open AccessEmerging Role of AI in Cardiovascular Medicine
This review provides an up-to-date overview of recent advances in artificial intelligence (AI) applied to cardiovascular medicine, spanning technological innovations, clinical applications, and broader societal implications. Over the past several years, deep learning methods have been increasingly adopted across multiple diagnostic modalities, including electrocardiography, echocardiography, coronary computed tomography angiography (CCTA), and intravascular ultrasound (IVUS), resulting in measurable improvements in diagnostic accuracy and workflow efficiency. In CCTA, AI facilitates automated detection of coronary stenosis and plaque burden. Echocardiography enables image segmentation and quantification of the left ventricular ejection fraction, while in IVUS, it is primarily applied to vessel structure analysis and lumen segmentation, and has also been used in conjunction with high-sensitivity troponin I to develop models capable of early ischemia assessment, offering the potential to rule out myocardial infarction from a single blood draw and to provide long-term risk estimates. More recently, domain-specific large language models, such as Med-PaLM, Gemini, and Articulate Medical Intelligence Explorer, have demonstrated diagnostic capabilities approaching physician-level performance in selected tasks. These systems are beginning to show promise for real-world clinical use by integrating multimodal data and providing interactive decision-making support. Nonetheless, several unresolved challenges persist, including the differences in cognitive reasoning between AI and clinicians, limited explainability, and complex regulatory and ethical considerations. Ongoing efforts, including randomized controlled trials (e.g., PROTEUS and RAPIDxAI) as well as the development of inclusive and sustainable digital health frameworks, aim to support the safe integration of AI into routine cardiovascular care. This review synthesizes current progress from technological, clinical, and implementation standpoints and explores the conditions under which AI can serve not as a replacement for clinicians, but as a collaborative partner in patient care.
- Open Access
Transcoder-based Circuit Analysis for Interpretable Single-Cell Foundation Models
Single-cell foundation models (scFMs) have demonstrated state-of-the-art performance on various tasks, such as cell-type annotation and perturbation response prediction, by learning gene regulatory networks from large-scale transcriptome data. However, a significant challenge remains: the decision-making processes of these models are less interpretable compared to traditional methods like differential gene expression analysis. Recently, transcoders have emerged as a promising approach for extracting interpretable decision circuits from large language models (LLMs). In this work, we train a transcoder on the cell2sentence (C2S) model, a state-of-the-art scFM. By leveraging the trained transcoder, we extract internal decision-making circuits from the C2S model. We demonstrate that the discovered circuits correspond to real-world biological mechanisms, confirming the potential of transcoders to uncover biologically plausible pathways within complex single-cell models.
Open AccessTranscoder-based Circuit Analysis for Interpretable Single-Cell Foundation Models
Single-cell foundation models (scFMs) have demonstrated state-of-the-art performance on various tasks, such as cell-type annotation and perturbation response prediction, by learning gene regulatory networks from large-scale transcriptome data. However, a significant challenge remains: the decision-making processes of these models are less interpretable compared to traditional methods like differential gene expression analysis. Recently, transcoders have emerged as a promising approach for extracting interpretable decision circuits from large language models (LLMs). In this work, we train a transcoder on the cell2sentence (C2S) model, a state-of-the-art scFM. By leveraging the trained transcoder, we extract internal decision-making circuits from the C2S model. We demonstrate that the discovered circuits correspond to real-world biological mechanisms, confirming the potential of transcoders to uncover biologically plausible pathways within complex single-cell models.
- Open Access
Application of Contrastive Learning on ECG Data: Evaluating Performance in Japanese and Classification with Around 100 Labels
The electrocardiogram (ECG) is a fundamental tool in cardiovascular diagnostics due to its powerful and non-invasive nature. One of the most critical usages is to determine whether more detailed examinations are necessary, with users ranging across various levels of expertise. Given this diversity in expertise, it is essential to assist users to avoid critical errors. Recent studies in machine learning have addressed this challenge by extracting valuable information from ECG data. Utilizing language models, these studies have implemented multimodal models aimed at classifying ECGs according to labeled terms. However, the number of classes was reduced, and it remains uncertain whether the technique is effective for languages other than English. To move towards practical application, we utilized ECG data from regular patients visiting hospitals in Japan, maintaining a large number of Japanese labels obtained from actual ECG readings. Using a contrastive learning framework, we found that even with 98 labels for classification, our Japanese-based language model achieves accuracy comparable to previous research. This study extends the applicability of multimodal machine learning frameworks to broader clinical studies and non-English languages.
Open AccessApplication of Contrastive Learning on ECG Data: Evaluating Performance in Japanese and Classification with Around 100 Labels
The electrocardiogram (ECG) is a fundamental tool in cardiovascular diagnostics due to its powerful and non-invasive nature. One of the most critical usages is to determine whether more detailed examinations are necessary, with users ranging across various levels of expertise. Given this diversity in expertise, it is essential to assist users to avoid critical errors. Recent studies in machine learning have addressed this challenge by extracting valuable information from ECG data. Utilizing language models, these studies have implemented multimodal models aimed at classifying ECGs according to labeled terms. However, the number of classes was reduced, and it remains uncertain whether the technique is effective for languages other than English. To move towards practical application, we utilized ECG data from regular patients visiting hospitals in Japan, maintaining a large number of Japanese labels obtained from actual ECG readings. Using a contrastive learning framework, we found that even with 98 labels for classification, our Japanese-based language model achieves accuracy comparable to previous research. This study extends the applicability of multimodal machine learning frameworks to broader clinical studies and non-English languages.