Hirotaka Ieki
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Hirotaka Ieki has a total of 147 co-authors. Here are the top co-authors by number of shared publications:
Andrew P. Ambrosy
7 shared publications- David Ouyang
Department of Cardiology, Smidt Heart Institute, Cedars-Sinai Medical Center, Los Angeles, California, USA
7 shared publications Bryan He
4 shared publicationsDavid Ouyang
4 shared publications- Issei Komuro
Department of Frontier Cardiovascular Science (I.K.), Graduate School of Medicine, the University of Tokyo, Japan
4 shared publications - Kaoru Ito
RIKEN Center for Integrative Medical Sciences
4 shared publications
Latest Publications
- Open Access
- 10 November 2026
- Published by Elsevier in Journal of Cardiac Failure
- Vol. 32 (11) , 1878-1888
Automated Deep Learning Pipeline for Characterizing Left Ventricular Diastolic Function
- 10 November 2026
- Published by Elsevier in Journal of Cardiac Failure
- Vol. 32 (11) , 1878-1888
Open AccessAutomated Deep Learning Pipeline for Characterizing Left Ventricular Diastolic Function
- Open Access
- 10 November 2026
- Published by Elsevier in Journal of Cardiac Failure
- Vol. 32 (11) , 2025-2028
Artificial Intelligence Detects Cardiac Amyloidosis Before Clinical Diagnosis
- 10 November 2026
- Published by Elsevier in Journal of Cardiac Failure
- Vol. 32 (11) , 2025-2028
Open AccessArtificial Intelligence Detects Cardiac Amyloidosis Before Clinical Diagnosis
- Open Access
Decoding personal and population variability in the human multiome
Background: With recent advances in technology, multiomic data have proliferated rapidly. Biobanks are now developing genomic, proteomic, lipidomic, and metabolomic data sets on the order of hundreds of thousands of people. However, bridging the gap between population-level data and the individual—a major challenge in precision medicine—remains a novel and challenging task. Methods: We analyzed the landscape of contemporary multiomics data and the drivers of population and personal molecular variability using a longitudinal cohort uniquely suited to precision health applications. Using healthy routine visit data (94 participants, median five visits per participant), we profiled the human multiome across eight molecular omic types and 1.2 million measurements, with replication in a sister multiomic cohort (322 participants). Results: Molecular data tended to cluster by omes, with the metabolome and proteome showing less intraomic correlation than other data types. While most blood-based molecular measurements followed log-normal distributions, between 2.0% and 8.1% of molecules showed evidence of bimodality, with cases driven by genetic polymorphisms or dietary differences (e.g., SERPINA5 [P = 1.4e-6], 2-piperidinone [P = 5.7e-5]). Molecular variation across the population was generally higher than that within an individual over time (e.g., xenobiotic metabolites, progestin steroids). Analyzing personal molecular variation identified several participants with notable temporal shifts (intraindividual CV > 4 SDs higher than population CV) in free fatty acids (FFA 20:1 and FFA 20:2) or cytokines (IL-17F and IFN-beta), which were unlikely to be detected cross-sectionally. Body mass index (BMI) had a much larger multiomic footprint (23.6% of molecules associated) than demographic factors like sex or age (5.7% and 1.9%, respectively), although machine-learning predictive models of sex (top ROC AUC of 0.98) were more successful than those for BMI (top R 2 of 0.59 by cytokines) or age (top R 2 of 0.34 by metabolites). Purely multiomic predictive models explained modest to high proportions of variation (R 2 = 0.22 to 0.94) in 20 common clinical tests, highlighting the potential clinical utility of these data even within a modestly sized cohort. Conclusion: Analysis of the molecular distributions, variation, and covariation of longitudinal multiomics data revealed clinically relevant differences at both the population and personal levels. Based on these findings, we conclude with six key considerations for the next generation of population omics as the field advances toward the goals of precision medicine. Multiomic data have proliferated rapidly in biomedical research, but many aspects of molecular variability remain poorly understood. We investigated in detail the drivers of personal and population variability in a longitudinal, multiomic cohort uniquely suited to precision health applications (94 participants, eight omic types, 1.2 million molecular measurements), with replication in a sister multiomic cohort (322 participants). While most molecules followed log-normal distributions, between 2.0% and 8.1% of molecules were bimodal, with cases driven by genetic polymorphisms or dietary differences (e.g., SERPINA5 [P = 1.4e-6], 2-piperidinone [P = 5.7e-5]). The metabolome and proteome showing less intraomic correlation than the lipidome or cytokines, and molecular variation across the population was generally higher than that within an individual over time with. Analyzing personal molecular variation identified several participants with notable temporal shifts (intraindividual CV > 4 SDs higher than population CV) in free fatty acids (FFA 20:1 and FFA 20:2) or cytokines (IL-17F and IFN-beta), which were unlikely to be detected cross-sectionally. Body mass index (BMI) was more broadly associated with molecular levels (23.6% of molecules associated) than were demographic factors like sex or age (5.7% and 1.9%, respectively), although machine-learning predictive models of sex (top ROC AUC of 0.98) were more successful than those for BMI (top R 2 of 0.59 by cytokines) or age (top R 2 of 0.34 by metabolites). Purely multiomic predictive models were modest to highly predictive (R 2 = 0.22 to 0.94) of 20 common clinical tests, highlighting the potential clinical utility of these data even within a modestly sized cohort. Our findings reveal six key considerations (missing variability, temporality, feasible clinicomics, complex modeling, aging complexity, and molecular matching) for precision medicine.
Open AccessDecoding personal and population variability in the human multiome
Background: With recent advances in technology, multiomic data have proliferated rapidly. Biobanks are now developing genomic, proteomic, lipidomic, and metabolomic data sets on the order of hundreds of thousands of people. However, bridging the gap between population-level data and the individual—a major challenge in precision medicine—remains a novel and challenging task. Methods: We analyzed the landscape of contemporary multiomics data and the drivers of population and personal molecular variability using a longitudinal cohort uniquely suited to precision health applications. Using healthy routine visit data (94 participants, median five visits per participant), we profiled the human multiome across eight molecular omic types and 1.2 million measurements, with replication in a sister multiomic cohort (322 participants). Results: Molecular data tended to cluster by omes, with the metabolome and proteome showing less intraomic correlation than other data types. While most blood-based molecular measurements followed log-normal distributions, between 2.0% and 8.1% of molecules showed evidence of bimodality, with cases driven by genetic polymorphisms or dietary differences (e.g., SERPINA5 [P = 1.4e-6], 2-piperidinone [P = 5.7e-5]). Molecular variation across the population was generally higher than that within an individual over time (e.g., xenobiotic metabolites, progestin steroids). Analyzing personal molecular variation identified several participants with notable temporal shifts (intraindividual CV > 4 SDs higher than population CV) in free fatty acids (FFA 20:1 and FFA 20:2) or cytokines (IL-17F and IFN-beta), which were unlikely to be detected cross-sectionally. Body mass index (BMI) had a much larger multiomic footprint (23.6% of molecules associated) than demographic factors like sex or age (5.7% and 1.9%, respectively), although machine-learning predictive models of sex (top ROC AUC of 0.98) were more successful than those for BMI (top R 2 of 0.59 by cytokines) or age (top R 2 of 0.34 by metabolites). Purely multiomic predictive models explained modest to high proportions of variation (R 2 = 0.22 to 0.94) in 20 common clinical tests, highlighting the potential clinical utility of these data even within a modestly sized cohort. Conclusion: Analysis of the molecular distributions, variation, and covariation of longitudinal multiomics data revealed clinically relevant differences at both the population and personal levels. Based on these findings, we conclude with six key considerations for the next generation of population omics as the field advances toward the goals of precision medicine. Multiomic data have proliferated rapidly in biomedical research, but many aspects of molecular variability remain poorly understood. We investigated in detail the drivers of personal and population variability in a longitudinal, multiomic cohort uniquely suited to precision health applications (94 participants, eight omic types, 1.2 million molecular measurements), with replication in a sister multiomic cohort (322 participants). While most molecules followed log-normal distributions, between 2.0% and 8.1% of molecules were bimodal, with cases driven by genetic polymorphisms or dietary differences (e.g., SERPINA5 [P = 1.4e-6], 2-piperidinone [P = 5.7e-5]). The metabolome and proteome showing less intraomic correlation than the lipidome or cytokines, and molecular variation across the population was generally higher than that within an individual over time with. Analyzing personal molecular variation identified several participants with notable temporal shifts (intraindividual CV > 4 SDs higher than population CV) in free fatty acids (FFA 20:1 and FFA 20:2) or cytokines (IL-17F and IFN-beta), which were unlikely to be detected cross-sectionally. Body mass index (BMI) was more broadly associated with molecular levels (23.6% of molecules associated) than were demographic factors like sex or age (5.7% and 1.9%, respectively), although machine-learning predictive models of sex (top ROC AUC of 0.98) were more successful than those for BMI (top R 2 of 0.59 by cytokines) or age (top R 2 of 0.34 by metabolites). Purely multiomic predictive models were modest to highly predictive (R 2 = 0.22 to 0.94) of 20 common clinical tests, highlighting the potential clinical utility of these data even within a modestly sized cohort. Our findings reveal six key considerations (missing variability, temporality, feasible clinicomics, complex modeling, aging complexity, and molecular matching) for precision medicine.
- 5 October 2026
- Published by Elsevier in JACC: Cardiovascular Imaging
- Vol. 19 (10) , 1186-1200
Comprehensive Aortic Stenosis Characterization Using Multiview Deep Learning
- 5 October 2026
- Published by Elsevier in JACC: Cardiovascular Imaging
- Vol. 19 (10) , 1186-1200
Comprehensive Aortic Stenosis Characterization Using Multiview Deep Learning
From task-specific AI to cardiovascular foundation models: a new era of multimodal interpretation
Artificial intelligence (AI) is reshaping cardiovascular medicine, evolving from task-specific supervised models towards multitask and foundation models capable of broader, more scalable interpretation. Early AI systems achieved high performance in electrocardiographic interpretation, echocardiographic measurement and disease diagnosis but their dependence on predefined outputs and labelled datasets limits flexibility and confines them to individual tasks. Foundation models address these limitations by learning reusable representations from large-scale data, often through self-supervised or contrastive pretraining, and adapting them to multiple downstream tasks. Emerging models in electrocardiography, echocardiography, chest radiography, cardiac MR and cardiac CT demonstrate capabilities including disease classification, quantitative measurement, segmentation, image-text retrieval, report generation and risk prediction. Generalist and multimodal biomedical foundation models further suggest a future in which cardiovascular AI integrates signals, imaging, clinical text, laboratory data and medical history to support patient-level diagnosis and decision-making. This review synthesises the transition from task-specific AI to cardiovascular foundation models and outlines opportunities for multimodal human-centred implementation.
From task-specific AI to cardiovascular foundation models: a new era of multimodal interpretation
Artificial intelligence (AI) is reshaping cardiovascular medicine, evolving from task-specific supervised models towards multitask and foundation models capable of broader, more scalable interpretation. Early AI systems achieved high performance in electrocardiographic interpretation, echocardiographic measurement and disease diagnosis but their dependence on predefined outputs and labelled datasets limits flexibility and confines them to individual tasks. Foundation models address these limitations by learning reusable representations from large-scale data, often through self-supervised or contrastive pretraining, and adapting them to multiple downstream tasks. Emerging models in electrocardiography, echocardiography, chest radiography, cardiac MR and cardiac CT demonstrate capabilities including disease classification, quantitative measurement, segmentation, image-text retrieval, report generation and risk prediction. Generalist and multimodal biomedical foundation models further suggest a future in which cardiovascular AI integrates signals, imaging, clinical text, laboratory data and medical history to support patient-level diagnosis and decision-making. This review synthesises the transition from task-specific AI to cardiovascular foundation models and outlines opportunities for multimodal human-centred implementation.
- Open Access
- 1 August 2026
- Published by Wolters Kluwer Health in Circulation: Genomic and Precision Medicine
- Vol. 19 (4) , e005341
Machine Learning Reveals the Contribution of Rare Genetic Variants and Enhances Risk Prediction for Coronary Artery Disease in the Japanese Population
BACKGROUND: GWASs (genome-wide association studies) have advanced our understanding of coronary artery disease (CAD) genetics and enabled the development of polygenic risk scores (PRSs) for estimating genetic risk based on common variant burden. However, GWASs have limitations in analyzing rare variants due to insufficient statistical power, thereby constraining PRS performance. METHODS: We conducted whole-genome sequencing of 1752 Japanese patients with CAD and 3019 controls. A machine learning-based analytical framework was applied to identify and interpret rare genetic variants associated with CAD pathogenesis. RESULTS: This approach identified 59 CAD-related genes, including known causal genes such as LDLR and those not previously captured by GWASs. A rare variant-based risk score derived from the framework demonstrated distinct clinical characteristics compared with a conventional common variant-based PRS. The rare variant-based risk score significantly discriminated CAD cases and predicted cardiovascular mortality in an independent cohort. Furthermore, combining the rare variant-based risk score with the traditional PRS improved CAD prediction compared with the PRS alone (area under the curve, 0.66 versus 0.61; P =0.007). CONCLUSIONS: These findings underscore the distinct and complementary value of the rare variant-based risk score compared with the conventional PRS, highlighting the enhanced predictive power achieved through their integration. This comprehensive approach proposes broader genetic profiling, offering substantial potential for improved clinical risk stratification and personalized prevention strategies.
- 1 August 2026
- Published by Wolters Kluwer Health in Circulation: Genomic and Precision Medicine
- Vol. 19 (4) , e005341
Open AccessMachine Learning Reveals the Contribution of Rare Genetic Variants and Enhances Risk Prediction for Coronary Artery Disease in the Japanese Population
BACKGROUND: GWASs (genome-wide association studies) have advanced our understanding of coronary artery disease (CAD) genetics and enabled the development of polygenic risk scores (PRSs) for estimating genetic risk based on common variant burden. However, GWASs have limitations in analyzing rare variants due to insufficient statistical power, thereby constraining PRS performance. METHODS: We conducted whole-genome sequencing of 1752 Japanese patients with CAD and 3019 controls. A machine learning-based analytical framework was applied to identify and interpret rare genetic variants associated with CAD pathogenesis. RESULTS: This approach identified 59 CAD-related genes, including known causal genes such as LDLR and those not previously captured by GWASs. A rare variant-based risk score derived from the framework demonstrated distinct clinical characteristics compared with a conventional common variant-based PRS. The rare variant-based risk score significantly discriminated CAD cases and predicted cardiovascular mortality in an independent cohort. Furthermore, combining the rare variant-based risk score with the traditional PRS improved CAD prediction compared with the PRS alone (area under the curve, 0.66 versus 0.61; P =0.007). CONCLUSIONS: These findings underscore the distinct and complementary value of the rare variant-based risk score compared with the conventional PRS, highlighting the enhanced predictive power achieved through their integration. This comprehensive approach proposes broader genetic profiling, offering substantial potential for improved clinical risk stratification and personalized prevention strategies.
- Open Access
Artificial Intelligence-Enabled Cardiac Function Estimation from Phone Videos of Echocardiograms
Importance: Mobile phone-recorded echocardiogram videos are commonly used in point-of-care, telemedicine, and resource-limited workflows, but artificial intelligence models for left ventricular ejection fraction (LVEF) estimation have primarily been evaluated on native Digital Imaging and Communications in Medicine (DICOM) videos. Objective: To evaluate whether previously described artificial intelligence models for LVEF estimation retain performance when applied to mobile phone-recorded echocardiographic videos. Design: Multicenter model validation study comparing model-estimated LVEF with clinician-reported LVEF. Setting: Three medical centers: Kaiser Permanente Northern California, Beth Israel Deaconess Medical Center through MIMIC-IV-ECHO, and Cedars-Sinai Medical Center. Participants: Source studies with clinician-reported LVEF and apical 4-chamber or apical 2-chamber views, yielding 6209 phone-recorded videos from 2648 studies and 2611 patients. Exposures: Mobile phone recording of native echocardiographic videos and fine-tuning of pretrained models using mobile phone-recorded videos from the Kaiser Permanente Northern California training cohort. Main Outcomes and Measures: Mean absolute error in ejection fraction percentage points, R² for continuous estimation, and area under the receiver operating characteristic curve for identifying ejection fraction greater than 50%. Results: The study included 6209 mobile phone-recorded echocardiographic videos from 2648 studies and 2611 patients; the weighted mean age was 68.4 years, and 1031 patients were male (39.5%). Without phone-video fine-tuning, the primary model achieved a mean absolute error of 7.00 percentage points, coefficient of determination of 0.49, and area under the receiver operating characteristic curve of 0.91 on phone-recorded videos; corresponding native DICOM performance was 6.08 percentage points, 0.60, and 0.93, respectively. On the 2396-video fine-tuning evaluation cohort, fine-tuning improved primary model performance to a mean absolute error of 6.96 percentage points, coefficient of determination of 0.61, and area under the receiver operating characteristic curve of 0.93. Fine-tuning the public EchoNet-Dynamic model improved performance from 9.36 percentage points, 0.37, and 0.84 to 7.86 percentage points, 0.50, and 0.89, respectively. Progressive central zoom preprocessing degraded model performance. Conclusions and Relevance: These findings suggest that artificial intelligence–assisted left ventricular ejection fraction estimation from mobile phone-recorded echocardiograms may be feasible when native image export is unavailable, although prospective evaluation is needed before clinical deployment. Key Points: Question: Can artificial intelligence models developed for native echocardiographic video formats estimate left ventricular ejection fraction from mobile phone-recorded echocardiogram videos? Findings: In this multicenter retrospective model-validation study of 6209 phone-recorded echocardiographic videos from 2648 studies, the primary pretrained model achieved a mean absolute error of 7.00 percentage points and area under the curve of 0.91 for identifying ejection fraction greater than 50%; phone-video fine-tuning improved performance and digital zoom degraded performance. Meaning: Artificial intelligence-assisted ejection fraction estimation from phone-recorded echocardiograms may support point-of-care and telemedicine workflows when native image export is unavailable.
Open AccessArtificial Intelligence-Enabled Cardiac Function Estimation from Phone Videos of Echocardiograms
Importance: Mobile phone-recorded echocardiogram videos are commonly used in point-of-care, telemedicine, and resource-limited workflows, but artificial intelligence models for left ventricular ejection fraction (LVEF) estimation have primarily been evaluated on native Digital Imaging and Communications in Medicine (DICOM) videos. Objective: To evaluate whether previously described artificial intelligence models for LVEF estimation retain performance when applied to mobile phone-recorded echocardiographic videos. Design: Multicenter model validation study comparing model-estimated LVEF with clinician-reported LVEF. Setting: Three medical centers: Kaiser Permanente Northern California, Beth Israel Deaconess Medical Center through MIMIC-IV-ECHO, and Cedars-Sinai Medical Center. Participants: Source studies with clinician-reported LVEF and apical 4-chamber or apical 2-chamber views, yielding 6209 phone-recorded videos from 2648 studies and 2611 patients. Exposures: Mobile phone recording of native echocardiographic videos and fine-tuning of pretrained models using mobile phone-recorded videos from the Kaiser Permanente Northern California training cohort. Main Outcomes and Measures: Mean absolute error in ejection fraction percentage points, R² for continuous estimation, and area under the receiver operating characteristic curve for identifying ejection fraction greater than 50%. Results: The study included 6209 mobile phone-recorded echocardiographic videos from 2648 studies and 2611 patients; the weighted mean age was 68.4 years, and 1031 patients were male (39.5%). Without phone-video fine-tuning, the primary model achieved a mean absolute error of 7.00 percentage points, coefficient of determination of 0.49, and area under the receiver operating characteristic curve of 0.91 on phone-recorded videos; corresponding native DICOM performance was 6.08 percentage points, 0.60, and 0.93, respectively. On the 2396-video fine-tuning evaluation cohort, fine-tuning improved primary model performance to a mean absolute error of 6.96 percentage points, coefficient of determination of 0.61, and area under the receiver operating characteristic curve of 0.93. Fine-tuning the public EchoNet-Dynamic model improved performance from 9.36 percentage points, 0.37, and 0.84 to 7.86 percentage points, 0.50, and 0.89, respectively. Progressive central zoom preprocessing degraded model performance. Conclusions and Relevance: These findings suggest that artificial intelligence–assisted left ventricular ejection fraction estimation from mobile phone-recorded echocardiograms may be feasible when native image export is unavailable, although prospective evaluation is needed before clinical deployment. Key Points: Question: Can artificial intelligence models developed for native echocardiographic video formats estimate left ventricular ejection fraction from mobile phone-recorded echocardiogram videos? Findings: In this multicenter retrospective model-validation study of 6209 phone-recorded echocardiographic videos from 2648 studies, the primary pretrained model achieved a mean absolute error of 7.00 percentage points and area under the curve of 0.91 for identifying ejection fraction greater than 50%; phone-video fine-tuning improved performance and digital zoom degraded performance. Meaning: Artificial intelligence-assisted ejection fraction estimation from phone-recorded echocardiograms may support point-of-care and telemedicine workflows when native image export is unavailable.
- Open Access
- 17 June 2026
- Published by Elsevier in JACC: Advances
- Vol. 5 (6) , 102801
Detection of Left Ventricular Outflow Obstruction From Standard B-Mode Echocardiogram Videos Using Deep Learning
- 17 June 2026
- Published by Elsevier in JACC: Advances
- Vol. 5 (6) , 102801
Open AccessDetection of Left Ventricular Outflow Obstruction From Standard B-Mode Echocardiogram Videos Using Deep Learning
- Open Access
Artificial Intelligence Detects Cardiac Amyloidosis Before Clinical Diagnosis
The approval of novel disease-modifying treatments for cardiac amyloidosis (CA) offers an avenue to stabilize disease progression. Timely diagnosis of CA is imperative so that patients can be connected to treatment earlier in their course. Echocardiography is a widely available, non-invasive modality to screen for CA. However, diagnosis of CA is often delayed due to phenotypic mimickers. Artificial intelligence (AI) applied to echocardiography may facilitate early detection and connect patients to more definitive diagnostic testing. In this case report, we analyzed the frequency of echocardiography prior to clinical diagnosis of amyloidosis and deployed an externally validated AI model in 349 patients with documented amyloidosis at Cedars-Sinai Medical Center. On average, AI detection was positive in patients with documented CA 218 days prior to clinical diagnosis and 209 days in all-comers. When integrated with clinical histories, AI may facilitate earlier detection of CA and connect patients to more timely management.
Open AccessArtificial Intelligence Detects Cardiac Amyloidosis Before Clinical Diagnosis
The approval of novel disease-modifying treatments for cardiac amyloidosis (CA) offers an avenue to stabilize disease progression. Timely diagnosis of CA is imperative so that patients can be connected to treatment earlier in their course. Echocardiography is a widely available, non-invasive modality to screen for CA. However, diagnosis of CA is often delayed due to phenotypic mimickers. Artificial intelligence (AI) applied to echocardiography may facilitate early detection and connect patients to more definitive diagnostic testing. In this case report, we analyzed the frequency of echocardiography prior to clinical diagnosis of amyloidosis and deployed an externally validated AI model in 349 patients with documented amyloidosis at Cedars-Sinai Medical Center. On average, AI detection was positive in patients with documented CA 218 days prior to clinical diagnosis and 209 days in all-comers. When integrated with clinical histories, AI may facilitate earlier detection of CA and connect patients to more timely management.
- Open Access
Automated Phenotyping of Mitral Stenosis Using Deep Learning
Background and Aims: Accurate classification of mitral stenosis (MS) remains a significant clinical challenge. This study aimed to develop an artificial intelligence (AI) framework to automatically detect clinically significant MS from echocardiography. Methods: We developed EchoNet-MS, an open-source end-to-end integrated approach combining video based convolutional neural networks to assess MS severity and differentiate rheumatic etiology from echocardiography and validated its performance across four cohorts. Results: EchoNet-MS was trained and validated in total of 431,612 videos from 44,671 studies from three different healthcare system. Combining assessments from multiple echocardiographic videos, the model was trained on a Kaiser Permanente Northern California (KPNC) cohort of 8,677 studies from 7,576 patients with a range of MS severity. The model was validated on a KPNC held-out test cohort (N=1,623) and a temporally distinct cohort (N=19,206), as well as Stanford Healthcare (SHC) cohort (N=3,333) and Cedars-Sinai Medical Center (CSMC) cohort (N=72,909). EchoNet-MS achieved excellent discrimination of severe MS with AUC 0.937 [95% CI: 0.913 – 0.958] in the KPNC held-out cohort, 0.994 [0.986 – 0.999] in the temporally distinct cohort, 0.991 [0.986 – 0.995] in SHC, and 0.973 [0.958 – 0.987] in CSMC. The model achieved excellent performance in classifying both rheumatic or non-rheumatic MS with AUC ranging from 0.890 and 0.967. Conclusions: EchoNet-MS accurately assesses MS severity and etiology using information from multiple echocardiographic views. Its strong performance generalizes robustly to external cohorts and shows potential as an automated clinical decision support tool.
Open AccessAutomated Phenotyping of Mitral Stenosis Using Deep Learning
Background and Aims: Accurate classification of mitral stenosis (MS) remains a significant clinical challenge. This study aimed to develop an artificial intelligence (AI) framework to automatically detect clinically significant MS from echocardiography. Methods: We developed EchoNet-MS, an open-source end-to-end integrated approach combining video based convolutional neural networks to assess MS severity and differentiate rheumatic etiology from echocardiography and validated its performance across four cohorts. Results: EchoNet-MS was trained and validated in total of 431,612 videos from 44,671 studies from three different healthcare system. Combining assessments from multiple echocardiographic videos, the model was trained on a Kaiser Permanente Northern California (KPNC) cohort of 8,677 studies from 7,576 patients with a range of MS severity. The model was validated on a KPNC held-out test cohort (N=1,623) and a temporally distinct cohort (N=19,206), as well as Stanford Healthcare (SHC) cohort (N=3,333) and Cedars-Sinai Medical Center (CSMC) cohort (N=72,909). EchoNet-MS achieved excellent discrimination of severe MS with AUC 0.937 [95% CI: 0.913 – 0.958] in the KPNC held-out cohort, 0.994 [0.986 – 0.999] in the temporally distinct cohort, 0.991 [0.986 – 0.995] in SHC, and 0.973 [0.958 – 0.987] in CSMC. The model achieved excellent performance in classifying both rheumatic or non-rheumatic MS with AUC ranging from 0.890 and 0.967. Conclusions: EchoNet-MS accurately assesses MS severity and etiology using information from multiple echocardiographic views. Its strong performance generalizes robustly to external cohorts and shows potential as an automated clinical decision support tool.