From: Machine learning techniques for arrhythmic risk stratification: a review of the literature
First author | Year of publication | Enrollment period | Number of patients | Age (years) | Males (%) | Type of clinical setting | Machine learning technique | Performance | Type of predicted arrhythmia | Features used for the prediction model |
---|---|---|---|---|---|---|---|---|---|---|
Hypertrophic cardiomyopathy | ||||||||||
Alis [24] | 2020 | 2014–2019 | 64 | 48.13 ± 13.06 | 65.6 | HCM | SVM, naïve Bayes, RF, k-NN | Accuracy SVM: 82.8% Naïve Bayes: 83.3% RF: 82.8% k-NN: 93.8% | Ventricular tachyarrhythmia | mean LGE% |
Bhattacharya [19] | 2019 | Mean duration: 2.86 years | 711 | No VA: 54 ± 15 Yes VA: 49 ± 16 | 61.0 | HCM | ensemble of logistic regression and natıve Bayes classifiers | Sens: 73%, spe: 76%, C-index = 0.83 | Sustained VT, VF | 22 features identified as predictive of VAr by the HCM-VAr-Risk Model |
Lyon [25] | 2018 | N/A | 123 | 47 ± 15 | 66.0 | HCM | Density-based clustering algorithm | N/A | VA | Data from 12 lead ECG holter |
Smole [20] | 2021 | N/A | 2302 | 46 ± 19 | 62.9 | HCM | RF, XGBoost, SVM, NN | Accuracy RF: 72% SVM (linear): 69% Bosted trees: 75% NN: 74% F1 score RF: 0.68 SVM (linear): 0.63 Boosted trees: 0.72 NN: 0,68 | VT | demographic characteristics, genetic data, clinical investigations, medications and disease-related events |
Ischemic heart disease—heart failure—cardiomyopathies | ||||||||||
Kotu [27] | 2015 | N/A | 54 | N/A | N/A | Post-myocardial infarction patients | K-NN, SVM, decision tree, RF | SVM classifier provided average accuracy of 92.6% and AUC of 0.921 | VA | Quantitative discriminative features extracted from LGE-CMR image |
Rogers [29] | 2021 | N/A | 42 | 64.7 ± 13.0 | 97.6 | Ischemic CMP | SVM, CΝΝ | Accuracy of the SVM: 83.2% Accuracy of convolutional NN: 56.7% | VT/VF | Ventricular monophasic action potentials |
Okada [28] | 2021 | 2003–2015 | 122 | 60 ± 11 | 87.0 | Ischemic CMP | SVM | Accuracy: 81% Correct classification: 86% NPP: 91% | VA | LGE-CMR data |
Au-yeung [67] | 2018 | NA | 788 | 60 | 77.3 | HF | RF, SVM | Both RF and SVM methods achieve a mean AUC of 0.81 for 5-min prediction and mean AUC of 0.87–0.88 for 10- second prediction | Ventricular tachyarrhythmia | VA onset prediction with heart rate variability data from ICD |
Meng [32] | 2019 | 2017–2019 | Retrospective: 500 Prospective: 1000 | N/A | N/A | HFrEF | Information gain ranking, decision trees, logistic regression, SVM, RF, ANN | NA (study protocol) | VT/VF | Demographic, clinical, biological, electrophysiological, social and psychological variables |
Wu [33] | 2020 | 2003–2015 | 382 | 57 ± 13 | 72.0 | HFrEF | RF | AUC: 0.88 | VA | clinical heart failure course, baseline CMR imaging metrics, levels of the interleukin-6 |
Rocon [37] | 2020 | 2011–2017 | 108 | 38.3 ± 15.5 | 48.1 | Non-compaction CMP | SVD impute, Parameter Selection Algorithm, sequential forward selection, distance-weighted k-NN | Accuracy: 75.5% Sens: 77% Spe: 75% | Major adverse cardiovascular events | LVEF (by CMRI), RV end systolic volume (by CMRI), RV systolic dysfunction (by echo) and RV lower diameter (by CMRI) |
Implantable devices | ||||||||||
Marzec [68] | 2018 | Minimum one year of observation | 235 | N/A | N/A | Patients with implantable electronic devices | Naïve, RF, decision tree analysis, k-NN, SVM | Accuracy Naïve: 74.5% RF: 76.6% Decision tree: 70.2% K-NN: 72.3% SVM: 74.5% F1 score Naïve: NA RF: 0.27 Decision tree: 0.13 K-NN: 0.00 SVM: NA | VT | Data about physical daily activities |
Shakibfar [35] | 2019 | N/A | 19,935 | N/A | N/A | ICD patients | RF | Accuracy: 96% AUC: 0.80 | VA (Electrical Storm) | ICD variables |
Channelopathies | ||||||||||
Lee [43] | 2021 | 1997–2017 | 516 | 50 ± 16 | 92.0 | Brugada syndrome | Nonnegative matrix factorisation, RF | Random survival forest Precision: 83.4% Recall: 85.3% F1 score: 84.3% Nonnegative matrix factorization Precision: 87.1% Recall: 88.8% F1 score: 87.9% | VT/VF, Brugada syndrome | Clinical and electrocardiographic data |
Congenital heart disease | ||||||||||
Dilller [51] | 2020 | 10 year observation period | 372 | 16.0 | 54.8 | Tetralogy of Fallot | Deep learning algorithms | N/A | VA prognosis, cardiac death, cardiac arrest | Prediction of adverse outcome using CMR data |
Sun [52] | 2021 | 2009 –2019 | 269 | N/A | 35.7 | Following ASD closure in pediatric patients | RF, SVM, K-nearest neighbor, AdaBoost, decision tree | Accuracy k-NN: 78.9% Decision tree: 82.6% AdaBoost: 84.95% SVM: 89.3% RF: 94.7% AUC k-NN: 0.82 Decision tree: 0.80 AdaBoost: 0.75 SVM: 0.87 RF: 0.90 | Postoperative arrhythmias | Demographic characteristics, cardiac imaging data and blood exams |
Different clinical settings | ||||||||||
Sessa [49] | 2020 | January 2015–December 2016 | 18,018 | 74.8 | 39.9 | Geriatric patients | Conditional inference tree | NA | VA | Specific medications |
Taye [58] | 2019 | N/A | N/A | N/A | N/A | Public databases | ANN classifiers | Accuracy: 98.6% Sens: 98.4% Spe: 99% | Ventricular tachyarrhythmia | QRS complex shape |
Okada [42] | 2019 | 2000–2017 | 76 | 53 ± 10 | 59.0 | Cardiac sarcoidosis | Supervised RF | Correct classification: 87% C‐statistic: 0.91 | VA, atrioventricular block | Regional scar burden |
Taye [57] | 2020 | N/A | 78 | 20.7–75.3 | 80.8 | Spontaneous ventricular tachyar- rhythmia database | CNN, k-NN, ANN, SVM | Accuracy CNN: 84.6% ANN: 73.5% SVM: 67.9% k-NN: 65.9% AUC CNN: 0.78 ANN: 0.65 SVM: 0.63 k-NN: 0.62 | Ventricular tachyarrhythmia | Heart rate variability signals |
Wu [54] | 2020 | 2012–2016 | 508 | 30.83 ± 6.17 | 75.0 | Young hypertensive patients | Recursive feature elimination, extreme gradient boosting | C statistic: 0.757 | Composite endpoint including sustained VT/VF | Demographics, medical history, vital signs, echocardiography, polysomnography, blood exams |
Bergau [47] | 2018 | N/A | N/A | N/A | N/A | N/A | SVM | Sens: 85% Spe: 90% | Torsades de pointes | Gene expression differences |
Chen [56] | 2021 | N/A | 17 | N/A | N/A | N/A | SVM, RF, XGboost | RF model with an average precision of 99.99% and recall of 88.98% | VA | Heartbeat interval time series |
Yap [48] | 2004 | N/A | N/A | N/A | N/A | N/A | SVM, probabilistic NN, k-NN, decision tree | Accuracy SVM: 97.4% Probabilistic NN: 71.8% k-NN: 89.7% Decision tree: 38.5% | Torsade de pointes | Set of agents |
Lee [55] | 2016 | 2013–2015 | N/A | N/A | N/A | N/A | ANN | Sens: 88.2% Spe: 82.4% PPV: 83.3% NPV: 87.5% | VT one hour prior to its occurrence | Heart rate variability and respiratory rate variability |