Table 4.
Example of ML for the Care of Subclinical and Clinical Stages of HF
| First Author (Ref. #) | Disease Application |
Sample Size |
Variable Input | Output | Algorithms | Results |
|---|---|---|---|---|---|---|
| Sabovcik et al (23) | Diagnosis of LVDD and LVH | 1,407 | 67 features and age, BMI, BP, history of hypertension, antihypertensive treatment, and electrocardiographic variables | Presence of LVDD and LVH | XGBoost, AdaBoost, RF, SVM, and LRM | The combination of ML and routinely measured data predicted LVDD and LVH with high accuracy. |
| Adedinsewo et al (60) | Diagnosis of LV systolic dysfunction | 1,606 | 12-lead ECG | Presence of reduced LVEF | CNN | CNN-enabled ECG algorithm effectively identified patients with LV systolic dysfunction, which outperformed NT-proBNP |
| Woolley et al (46) | Classification of HFpEF | 429 | 363 biomarkers | HFpEF subgroups | Cluster analysis | Cluster analysis identified four subgroups of HFpEF patients with distinct biomarker profiles, clinical features and outcomes |
| Narang et al (66) | Automated imaging acquisition | 240 | Echocardiographic indices | Cardiac view and echocardiographic indices | CNN | CNN-enabled automated image acquisition algorithm allowed novices in ultrasonography to obtain cardiac views for cardiac structure/function evaluation |
| Knackstedt et al (61) | Automated assessment of LVEF and LS | 255 | Apical 4- and 2-chamber echocardiographic imaging | Measures of LVEF and LS | ML-enabled software (AutoLV) | ML-enabled software provided rapid and reproducible assessment of LVEF and LS |
| Zhang et al (62) | Automated image interpretation | 14,035 Echocardiographic imaging data | Echocardiographic imaging | Cardiac view identification, chamber segmentation, and cardiac structure and function metrics | CNN-enabled fully automated assessment | CNN-enabled fully automated assessment identified cardiac view, segmented cardiac chamber, measured cardiac structure and function with high accuracy |
| Frizzell et al (92) | Prediction of readmission | 56,477 | Demographics, socioeconomics, medical history, HF characterization, medications used, vital signs, body weights, laboratories, and discharge interventions | 30-day HF readmission | Tree-augmented naïve Bayesian network, LRM with backward stepwise selection, LRM with LASSO, gradient boosted model, and RF | Prediction of 30-day HF readmissions was similar between ML models and traditional prediction models. |
| Ahmad et al (45) | Prediction of clinical outcome and HF subgroups classification | 44,886 | 86 variables for RF model; variables for cluster analysis include age, heart rate, creatinine, hemoglobin, weight, systolic BP, mean arterial pressure, and income | One-year survival; HF classification | RF and cluster analysis | ML models accurately predicted outcomes for HF patients. Cluster analysis identified 4 distinct HF phenotypes that differed significantly in outcomes and in response to therapy |
BMI = body mass index; CNN = convolutional neural network; HFpEF = heart failure with preserved ejection fraction; LASSO = Least Absolute Shrinkage and Selection Operator; LS = longitudinal strain; LVDD = left ventricular diastolic dysfunction; LVH = left ventricular hypertrophy; NT-proBNP = N-terminal pro-B type natriuretic peptide; SVM = support vector machine; other abbreviations as in Tables 1 and 3.