Table 1.
Study characteristics and predictive performance of studies included studies reporting on a machine learning and deep learning model.
| Study characteristics |
ML and DL modelling |
Performance |
Validation |
|||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Author | n | Subjects | Study design | Endpoint (Prevalence) | Features | Algorithm | Sensitivity | Specificity | PPV | NPV | AUC | Accuracy | Internal validation | External validation |
| Au-Yeung et al.15 | 788 | Prophylactic ICD recipient (77.3% male) | RCT | Appropriate ICD shock (3.3%) | HRV (non-linear domain, frequency-domain) | SVM and RF∗ | 74.0 | 74.0 | N/A | N/A | 81.0 | N/A | 80% training, 20% test | N/A |
| Do et al.38 | 1874 | Hospitalised (66.9% male) | Retrospective, case–control study | IHCA (5.1%) | Trend analysis (slope, change) | RF∗, LR | 94.6 | 63.2 | N/A | N/A | 82.9 | N/A | 80% train, 20% test | N/A |
| Lee et al.53 | 82 (104 recordings) | Hospitalised (sex distribution unknown) | Prospective cohort | VT (50%) | HRV (time-domain, non-linear Poincare, frequency-domain) | ANN | 70.6 | 76.5 | 75.0 | 72.2 | 75.0 | N/A | 60% train, 40% test | N/A |
| Kwon et al.16 | 25 672 | Hospitalised (53.1% male) | Retrospective cohort | IHCA (2.07%) | N/A | CNN | 77.8 | 92.0 | 76.0 | 99.8 | 94.8 | N/A | 70% train, 30% test | Yes (n = 10,728) |
| Gleeson et al.45 | 295 | Prophylactic ICD (74.2% male) | Retrospective cohort | ICD implantation or mortality (16.6%) | Spatial ECG parameters, complexity parameters and conventional ECG parameters | DT | N/A | N/A | N/A | N/A | 75.0 | N/A | 60%, 40% test | N/A |
| Martinez-Alanis et al.56 | 91 | ICD carriers (93.4% male) | Prospective cohort study | SCD (50%) | HRV (frequency and time-domain) and Heartprint Indices | SVM | N/A | N/A | N/A | N/A | 68.0 | 67.65 | 10-fold CV | Yes |
| Ong et al.17 | 925 | ED admissions (61.9% male) | Prospective cohort study | IHCA (4.6%) | HRV (time-domain, frequency-domain, and geometric parameters.) | SVM | 81.4 | 72.3 | 12.5 | 98.8 | 78.1 | N/A | LOOCV | N/A |
| Ramirez et al.62 | 597 | CHF (71.2% male) | Prospective cohort study | SCD (8.2%) | ECG risk makers (repolarisation dispersion, TWA, HRT) | SVM | 18.0 | 79.0 | N/A | N/A | N/A | N/A | 5-fold CV | N/A |
| Rodriguez et al.64 | 91 | Idiopathic dilated cardiomyopathy (sex distribution unknown) | Prospective cohort study | VT/VF or SCD (15.4%) | HRV (time-domain, frequency-domain and non-linear Poincaré) | SVM | 92.9 | 98.0 | N/A | N/A | 95.0 | 96.8 | LOOCV | N/A |
| Rogers et al.18 | 42 | Ischaemic cardiomyopathy (97.8% male) | Prospective cohort study | VT/VF (30.9%) | Mathematical timeserie features | SVM∗, CNN | 84.6 | 86.2 | 73.3 | 92.6 | 90.0 | 85.7 | 70% training, 30% testing | N/A |
ANN = artificial neural network, AUC = area under the curve, CNN = convolutional neural network, CHF = congestive heart failure, CV = cross validation, DT = decision tree, ECG = electrocardiography, ED = emergency department, RF = random forest, LOOCV = leave-one-out cross validation, LR = logistic regression, HRT = heart rate turbulence, HRV = heart rate variability, IHCA = in-hospital cardiac arrest, ICD = implantable cardioverter defibrillator, LOOCV = leave-one-out cross validation, N/A = not applicable, NPV = negative predictive value, PPV = positive predictive value, RCT = randomised controlled trial, SCD = sudden cardiac death, SVM = support vector machine, TWA = T-wave alternans, VT = ventricular tachycardia, VF = ventricular fibrillation.