User login
A machine-learning model that uses readily available clinical and electrocardiography data may have the potential to identify left ventricular (LV) diastolic dysfunction, a key biomarker in predicting heart failure, without echocardiography, but a workable clinical platform is still far off, a team of North American researchers reported.
“This cost-effective strategy may be a valuable first clinical step for assessing the presence of LV dysfunction and may potentially aid in the early diagnosis and management of heart failure patients,” Nobuyuki Kagiyama, MD, PhD, of West Virginia University, Morgantown, and colleagues, wrote in the Journal of the American Academy of Cardiology.
The researchers reported on a multicenter, prospective study that evaluated 1,202 patients from three centers in the United States and one in Canada. To develop machine-learning models, the study pooled 814 patients from the U.S. institutions as an internal cohort. They were then randomly divided into a training set and an internal test set on an 80:20 basis (651 and 163). The 388 Canadian patients were reserved as an external set to test the model.
All patients had 12-lead ECG and simultaneous body surface signal-processed ECG (spECG) along with comprehensive two-dimensional Doppler ECG on the same day.
How the model works
The machine-learning model estimated echocardiographic LV relaxation velocities (e’) values using traditional ECG and spECG features. The model also took into account 10 basic clinical features: age; sex; systolic and diastolic blood pressure; and comorbid conditions such as cerebrovascular and cardiovascular disease, diabetes, hypertension, dyslipidemia, and chronic kidney disease.
Patient characteristics were starkly different between the internal (United States) and external (Canadian) cohorts, with the latter being 10 years older on average (65 vs. 44; P < .001), predominantly male (58.2% vs. 47.3%; P < .001) and with significantly lower rates of coronary artery disease (1.8% vs. 21.1%; P < .001), although average blood pressure was similar between the two groups.
The study used area under the curve (AUC) to calculate the predictability of the machine-learning estimated e’ values versus the guideline-based reduced e’, finding close correlation between the internal (AUC, 0.83; sensitivity, 78%; specificity, 77%; negative predictive value, 73%; and positive predictive value, 82%) and external test sets (AUC, 0.84; sensitivity, 90%; specificity, 61%; NPV, 81%; and PPV, 77%).
Similar variations between the two cohorts were reported for global LV diastolic dysfunction and reduced LV ejection fraction.
The final model used 18 features in all, including 3 clinical features (age, dyslipidemia, and hypertension), 7 scores from spECG features, and 8 from traditional ECG features.
Interpreting the results
Dr. Kagiyama and colleagues noted that, because impaired myocardial relaxation is an early sign of cardiac tissue deterioration, screening for it can aid in early detection of subclinical LVDD and earlier treatment for hypertension and diabetes. But they acknowledged that further studies are needed.
In an invited editorial, Khurram Nasir, MD, MPH, MSc, of Houston Methodist DeBakey Heart and Vascular Center and Rohan Khera, MD, MS, of Yale University, New Haven, Conn., wrote that the machine-learning model has a way to go.
They noted that the 73%-77% accuracy of the model in identifying diastolic dysfunction impedes its imminent use. “Although we are excited about the prospects of such developments, we hold out for better evidence for their actual use,” they wrote, adding that the algorithms have limited use in the clinic because most patients already get “definitive testing” if they need it.
Developing a machine-learning model that obviates the need for ECG for evaluating LV diastolic dysfunction seems dubious at this time, said Luigi Di Biase, MD, PhD, section head of electrophysiology and director of arrhythmia services at Montefiore Medical Center and professor at Albert Einstein College of Medicine, both in New York. “The echo is not a difficult test. It’s the most proven usable tool that we have in cardiology because it’s easy to reproduce, low cost, and noninvasive – so we have all that we want in medicine.”
But machine learning does have potential, added Dr. Di Biase, who’s also a member of the American College of Cardiology’s Electrophysiology Section Leadership Council. “If this application could predict the people that would develop diastolic dysfunction that leads to heart failure – because an echo at that time may be negative but there may be other features that tell me this patient will develop disease – then it would have a much different clinical impact.”
The National Science Foundation provided funding for the study. Heart Test Laboratories, doing business as Heart Sciences, provided funding and spECG devices. Dr. Kagiyama reported receiving a research grant from Hitachi Healthcare. A coauthor disclosed financial relationships with Heart Sciences, Ultronics, and Kencor Health.
Dr. Nasir, Dr. Khera, and Dr. Di Biase have no relevant financial relationships to disclose.
SOURCE: Kagiyama N et al. J Am Coll Cardiol. 2020;76:930-41.
A machine-learning model that uses readily available clinical and electrocardiography data may have the potential to identify left ventricular (LV) diastolic dysfunction, a key biomarker in predicting heart failure, without echocardiography, but a workable clinical platform is still far off, a team of North American researchers reported.
“This cost-effective strategy may be a valuable first clinical step for assessing the presence of LV dysfunction and may potentially aid in the early diagnosis and management of heart failure patients,” Nobuyuki Kagiyama, MD, PhD, of West Virginia University, Morgantown, and colleagues, wrote in the Journal of the American Academy of Cardiology.
The researchers reported on a multicenter, prospective study that evaluated 1,202 patients from three centers in the United States and one in Canada. To develop machine-learning models, the study pooled 814 patients from the U.S. institutions as an internal cohort. They were then randomly divided into a training set and an internal test set on an 80:20 basis (651 and 163). The 388 Canadian patients were reserved as an external set to test the model.
All patients had 12-lead ECG and simultaneous body surface signal-processed ECG (spECG) along with comprehensive two-dimensional Doppler ECG on the same day.
How the model works
The machine-learning model estimated echocardiographic LV relaxation velocities (e’) values using traditional ECG and spECG features. The model also took into account 10 basic clinical features: age; sex; systolic and diastolic blood pressure; and comorbid conditions such as cerebrovascular and cardiovascular disease, diabetes, hypertension, dyslipidemia, and chronic kidney disease.
Patient characteristics were starkly different between the internal (United States) and external (Canadian) cohorts, with the latter being 10 years older on average (65 vs. 44; P < .001), predominantly male (58.2% vs. 47.3%; P < .001) and with significantly lower rates of coronary artery disease (1.8% vs. 21.1%; P < .001), although average blood pressure was similar between the two groups.
The study used area under the curve (AUC) to calculate the predictability of the machine-learning estimated e’ values versus the guideline-based reduced e’, finding close correlation between the internal (AUC, 0.83; sensitivity, 78%; specificity, 77%; negative predictive value, 73%; and positive predictive value, 82%) and external test sets (AUC, 0.84; sensitivity, 90%; specificity, 61%; NPV, 81%; and PPV, 77%).
Similar variations between the two cohorts were reported for global LV diastolic dysfunction and reduced LV ejection fraction.
The final model used 18 features in all, including 3 clinical features (age, dyslipidemia, and hypertension), 7 scores from spECG features, and 8 from traditional ECG features.
Interpreting the results
Dr. Kagiyama and colleagues noted that, because impaired myocardial relaxation is an early sign of cardiac tissue deterioration, screening for it can aid in early detection of subclinical LVDD and earlier treatment for hypertension and diabetes. But they acknowledged that further studies are needed.
In an invited editorial, Khurram Nasir, MD, MPH, MSc, of Houston Methodist DeBakey Heart and Vascular Center and Rohan Khera, MD, MS, of Yale University, New Haven, Conn., wrote that the machine-learning model has a way to go.
They noted that the 73%-77% accuracy of the model in identifying diastolic dysfunction impedes its imminent use. “Although we are excited about the prospects of such developments, we hold out for better evidence for their actual use,” they wrote, adding that the algorithms have limited use in the clinic because most patients already get “definitive testing” if they need it.
Developing a machine-learning model that obviates the need for ECG for evaluating LV diastolic dysfunction seems dubious at this time, said Luigi Di Biase, MD, PhD, section head of electrophysiology and director of arrhythmia services at Montefiore Medical Center and professor at Albert Einstein College of Medicine, both in New York. “The echo is not a difficult test. It’s the most proven usable tool that we have in cardiology because it’s easy to reproduce, low cost, and noninvasive – so we have all that we want in medicine.”
But machine learning does have potential, added Dr. Di Biase, who’s also a member of the American College of Cardiology’s Electrophysiology Section Leadership Council. “If this application could predict the people that would develop diastolic dysfunction that leads to heart failure – because an echo at that time may be negative but there may be other features that tell me this patient will develop disease – then it would have a much different clinical impact.”
The National Science Foundation provided funding for the study. Heart Test Laboratories, doing business as Heart Sciences, provided funding and spECG devices. Dr. Kagiyama reported receiving a research grant from Hitachi Healthcare. A coauthor disclosed financial relationships with Heart Sciences, Ultronics, and Kencor Health.
Dr. Nasir, Dr. Khera, and Dr. Di Biase have no relevant financial relationships to disclose.
SOURCE: Kagiyama N et al. J Am Coll Cardiol. 2020;76:930-41.
A machine-learning model that uses readily available clinical and electrocardiography data may have the potential to identify left ventricular (LV) diastolic dysfunction, a key biomarker in predicting heart failure, without echocardiography, but a workable clinical platform is still far off, a team of North American researchers reported.
“This cost-effective strategy may be a valuable first clinical step for assessing the presence of LV dysfunction and may potentially aid in the early diagnosis and management of heart failure patients,” Nobuyuki Kagiyama, MD, PhD, of West Virginia University, Morgantown, and colleagues, wrote in the Journal of the American Academy of Cardiology.
The researchers reported on a multicenter, prospective study that evaluated 1,202 patients from three centers in the United States and one in Canada. To develop machine-learning models, the study pooled 814 patients from the U.S. institutions as an internal cohort. They were then randomly divided into a training set and an internal test set on an 80:20 basis (651 and 163). The 388 Canadian patients were reserved as an external set to test the model.
All patients had 12-lead ECG and simultaneous body surface signal-processed ECG (spECG) along with comprehensive two-dimensional Doppler ECG on the same day.
How the model works
The machine-learning model estimated echocardiographic LV relaxation velocities (e’) values using traditional ECG and spECG features. The model also took into account 10 basic clinical features: age; sex; systolic and diastolic blood pressure; and comorbid conditions such as cerebrovascular and cardiovascular disease, diabetes, hypertension, dyslipidemia, and chronic kidney disease.
Patient characteristics were starkly different between the internal (United States) and external (Canadian) cohorts, with the latter being 10 years older on average (65 vs. 44; P < .001), predominantly male (58.2% vs. 47.3%; P < .001) and with significantly lower rates of coronary artery disease (1.8% vs. 21.1%; P < .001), although average blood pressure was similar between the two groups.
The study used area under the curve (AUC) to calculate the predictability of the machine-learning estimated e’ values versus the guideline-based reduced e’, finding close correlation between the internal (AUC, 0.83; sensitivity, 78%; specificity, 77%; negative predictive value, 73%; and positive predictive value, 82%) and external test sets (AUC, 0.84; sensitivity, 90%; specificity, 61%; NPV, 81%; and PPV, 77%).
Similar variations between the two cohorts were reported for global LV diastolic dysfunction and reduced LV ejection fraction.
The final model used 18 features in all, including 3 clinical features (age, dyslipidemia, and hypertension), 7 scores from spECG features, and 8 from traditional ECG features.
Interpreting the results
Dr. Kagiyama and colleagues noted that, because impaired myocardial relaxation is an early sign of cardiac tissue deterioration, screening for it can aid in early detection of subclinical LVDD and earlier treatment for hypertension and diabetes. But they acknowledged that further studies are needed.
In an invited editorial, Khurram Nasir, MD, MPH, MSc, of Houston Methodist DeBakey Heart and Vascular Center and Rohan Khera, MD, MS, of Yale University, New Haven, Conn., wrote that the machine-learning model has a way to go.
They noted that the 73%-77% accuracy of the model in identifying diastolic dysfunction impedes its imminent use. “Although we are excited about the prospects of such developments, we hold out for better evidence for their actual use,” they wrote, adding that the algorithms have limited use in the clinic because most patients already get “definitive testing” if they need it.
Developing a machine-learning model that obviates the need for ECG for evaluating LV diastolic dysfunction seems dubious at this time, said Luigi Di Biase, MD, PhD, section head of electrophysiology and director of arrhythmia services at Montefiore Medical Center and professor at Albert Einstein College of Medicine, both in New York. “The echo is not a difficult test. It’s the most proven usable tool that we have in cardiology because it’s easy to reproduce, low cost, and noninvasive – so we have all that we want in medicine.”
But machine learning does have potential, added Dr. Di Biase, who’s also a member of the American College of Cardiology’s Electrophysiology Section Leadership Council. “If this application could predict the people that would develop diastolic dysfunction that leads to heart failure – because an echo at that time may be negative but there may be other features that tell me this patient will develop disease – then it would have a much different clinical impact.”
The National Science Foundation provided funding for the study. Heart Test Laboratories, doing business as Heart Sciences, provided funding and spECG devices. Dr. Kagiyama reported receiving a research grant from Hitachi Healthcare. A coauthor disclosed financial relationships with Heart Sciences, Ultronics, and Kencor Health.
Dr. Nasir, Dr. Khera, and Dr. Di Biase have no relevant financial relationships to disclose.
SOURCE: Kagiyama N et al. J Am Coll Cardiol. 2020;76:930-41.
FROM THE JOURNAL OF THE AMERICAN COLLEGE OF CARDIOLOGY