Log in to MyACC
Machine Learning for Echo Assessment of Diastolic Dysfunction
Quick Takes
- In the deep neural network model, e’ (LV early diastolic relaxation velocity) was the most important echocardiographic variable for risk prediction, followed by E/e’ (early diastolic transmitral flow velocity/e’), LV mass, EF, and left atrial size.
- The model reclassified 73% of patients with grade I (mild) diastolic dysfunction and 80% of those with indeterminate diastolic function into the high-risk phenogroup.
Study Questions:
Can a deep neural network (DeepNN) model, integrating multidimensional echocardiographic data, enhance risk stratification for patients with heart failure with preserved ejection fraction (HFpEF)?
Methods:
An unsupervised machine learning approach involving similarity clustering was used to assign patients in a training cohort to low- and high-risk groups, based on nine echocardiographic variables: left ventricular ejection fraction (LVEF), LV mass index, early diastolic transmitral flow velocity (E), late diastolic transmitral flow velocity (A), E/A ratio, early diastolic relaxation velocity (e’), E/e’ ratio, left atrial volume index, and tricuspid regurgitation peak velocity. A cloud-based automated machine learning platform, with supervision, was then used to create a DeepNN for predicting low- and high-risk phenogroups of LV diastolic dysfunction. Two external validation cohorts were used to assess the model’s prediction of LV filling pressure among patients who underwent right or left heart catheterization and clinical outcomes among patients with varying degrees of LV systolic and diastolic dysfunction. Subsequently, data from three HFpEF trials (TOPCAT echocardiography substudy, RELAX-HF, and NEAT-HFpEF) were used to assess the model’s prediction of long-term clinical outcomes.
Results:
In the model development cohort (n = 990), e’ was the most important echocardiographic variable for risk prediction, followed by E/e’, LV mass, EF, and left atrial size. In the clinical outcome external validation cohort (n = 219), high-risk patients were older (mean age 65 vs. 50 years, p < 0.001) and were more likely to have diabetes, hypertension, and atrial fibrillation. In the hemodynamic validation cohort (n = 84), high-risk patients had higher LV filling pressure (p = 0.004). Compared with echocardiographic guideline-based assessment of diastolic function, the DeepNN model better predicted elevated LV filling pressure (area under the curve 0.88 vs. 0.67, p = 0.01).
In the TOPCAT cohort (n = 518), the high-risk phenogroup had higher rates of heart failure hospitalization or cardiac death as compared with the low-risk phenogroup (hazard ratio [HR], 1.92; 95% confidence interval [CI], 1.16-3.32; p = 0.01). Spironolactone was associated with a lower risk of heart failure hospitalization or cardiac death in the high-risk phenogroup (HR, 0.65; 95% CI, 0.46-0.90; p = 0.01), but not in the low-risk phenogroup. While 94% of patients with grade II or III (advanced) diastolic dysfunction on guideline-based assessment fell into the high-risk phenogroup, the DeepNN model reclassified 73% of patients with grade I (mild) diastolic dysfunction and 80% of those with indeterminate diastolic function into the high-risk phenogroup.
In the pooled RELAX-HF/NEAT-HFpEF cohort (n = 346), patients in the high-risk phenogroup had higher troponin (p < 0.001) and NTpro–B-type natriuretic peptide (p < 0.001) levels, and lower peak oxygen consumption on exercise testing (p = 0.001).
Conclusions:
This DeepNN model enhances traditional echocardiography-based risk assessment for patients with HFpEF and may identify a subgroup of patients who can benefit most from spironolactone therapy.
Perspective:
The greatest utility of this model is likely its ability to risk reclassify patients with mild or indeterminate diastolic dysfunction based on standard echocardiographic assessment. Among patients with HFpEF, common comorbidities such as lung disease, obesity, and atrial fibrillation may limit echocardiographic image quality and quantification of diastolic function parameters. Therefore, real-world validation of this model will be important before its use can be recommended in clinical practice.
Clinical Topics: Arrhythmias and Clinical EP, Heart Failure and Cardiomyopathies, Noninvasive Imaging, Prevention, Atrial Fibrillation/Supraventricular Arrhythmias, Acute Heart Failure, Heart Failure and Cardiac Biomarkers, Echocardiography/Ultrasound, Hypertension
Keywords: Artificial Intelligence, Atrial Fibrillation, Cardiac Catheterization, Diabetes Mellitus, Diastole, Echocardiography, Diagnostic Imaging, Exercise Test, Heart Failure, Hypertension, Lung Diseases, Natriuretic Peptide, Brain, Obesity, Oxygen Consumption, Risk Assessment, Secondary Prevention, Spironolactone, Stroke Volume, Tricuspid Valve Insufficiency, Troponin, Ventricular Function, Left
< Back to Listings
