Table 2.
Previous studies regarding the application of AI models in cardiomyopathy. ARVC: arrhythmogenic right ventricular cardiomyopathy; CCN: convolutional neural network; DCM: dilated cardiomyopathy; DT: decision tree; HCM: hypertrophic cardiomyopathy; LR: logistic regression; K-NN: K-nearest neighbor; NDLVC: non-dilated left ventricular cardiomyopathy; RCM: restrictive cardiomyopathy; RF: random forest; SVM: support vector machine; TTE: transthoracic echocardiogram; XGboost: extreme gradient boosting.
| Authors | Year | Patients | Cardiomyopathy Phenotypes |
Variables | AI Models | Programming Languages | Model Interpretability | Validation | Public Datasets and Code | Results |
|---|---|---|---|---|---|---|---|---|---|---|
| Haimovich et al. [38] |
2023 | 50,709 | HCM | ECG data | CNN | Python v3.8 (Python Software Foundation, Beaverton, Oregon) |
Black box | External validation | No public datasets No algorithm code available |
The DL model achieved an AUROC of 0.95 [95% CI, 0.93–0.97] for cardiac amyloidosis, 0.92 [95% CI, 0.90–0.94] for hypertrophic cardiomyopathy, 0.90 [95% CI, 0.88–0.92] for aortic stenosis, 0.76 [95% CI, 0.76–0.77] for hypertensive left ventricle hypertrophy, and 0.69 [95% CI 0.68–0.71] for other left ventricle hypertrophy etiologies. |
| Beneyto et al. [39]. |
2023 | 591 | HCM | Clinical, laboratory and TTE data | DT, RF, and SVM | R packages version 4.1.1 (R Foundation for Statistical Computing, Vienna, Austria) |
Inherently interpretable model Post-modeling explainability |
Internal validation | No public datasets No algorithm code available |
The proposed ML model achieved an AUROC of 0.90 (0.85–0.94), sensitivity of 0.31 (0.17–0.44), specificity of 0.96 (0.91–0.99), and accuracy of 0.80 (0.75–0.85) in the testing set. |
| Siontis et al. [40] |
2021 | 18,739 | HCM | ECG data | CNN | Python (Python Software Foundation, Beaverton, Oregon) |
Black box | External validation | No public datasets No algorithm code available |
The DL model achieved an AUC of 0.98 (95% CI 0.98–0.99), sensitivity of 92%, specificity of 95%, positive predictive value of 22%, and negative predictive value of 99%. |
| Hwang et al. [41] |
2022 | 930 | HCM | TTE data | CNN | R packages version 4.1.1 (R Foundation for Statistical Computing, Vienna, Austria) |
Black box | Internal validation | No public datasets Algorithm code available (https://github.com/djchoi1742/Echo_LVH) accessed on 30 December 2023 |
The DL model achieved average AUCs of 0.962, 0.982, and 0.996 in the test sets for hypertensive heart disease, HCM, and cardiac amyloidosis, respectively. |
| Baeßler et al. [42] |
2018 | 32 | HCM | Radiomics features | Machine learning | R packages Version 3.4.0 (R Foundation for Statistical Computing, Vienna, Austria) |
Post-modeling explainability | Internal validation | No public datasets No algorithm code available |
The proposed ML-based model achieved an AUC of 0.95, with a diagnostic sensitivity of 91% and a specificity of 93%. |
| Zhang et al. [43] |
2023 | 238 | HCM, DCM | Radiomics features | Multilayer perceptron, DT, RF, LR, XGboost, SVM, naive Bayes, K-nearest neighbor, and ensemble learning | Python version 4.1.1 (Python Software Foundation, Beaverton, Oregon) |
Inherently interpretable model Post-modeling explainability |
Internal validation | Public datasets (https://acdc.creatis.insalyon.fr/description/databases.html; https://www.ub.edu/mnms/) accessed on 30 December 2023 No algorithm code available |
The proposed ML model achieved an accuracy of 91.2% and average AUCs of 0.962, 0.982, and 0.996 in the test sets for HCM, DCM, and healthy controls, respectively. |
| Tayal et al. [44] |
2022 | 665 | DCM | Demographic, clinical, genetic, CMR, and proteomics parameters | RF | Premium R packages (R Foundation for Statistical Computing, Vienna, Austria) |
Post-modeling explainability | External validation | No public datasets No algorithm code available |
The proposed ML model identified three novel DCM subtypes: profibrotic metabolic, mild nonfibrotic, and biventricular impairment. |
| Zhou et al. [45] |
2023 | 399 | DCM | Clinical and TTE data | RF, LR, neural network, and XGBoost | R packages version 3.6.2 (R Foundation for Statistical Computing, Vienna, Austria) |
Inherently interpretable model Post-modeling explainability |
External validation | No public datasets No algorithm code available |
The ML model achieved good accuracy in discriminating between different etiologies in an external validation cohort with a sensitivity of 64%, a specificity of 93%, and AUC of 0.804. |
| Shrivastava et al. [46] |
2021 | 16,471 | DCM | ECG data | CNN | Python version 3.9 (Python Software Foundation, Beaverton, Oregon) |
Black box | Internal validation | No public datasets No algorithm code available |
The diagnostic performance of the proposed DL models yielded an AUC of 0.955, a sensitivity of 98.8%, and specificity of 44.8%, a negative predictive value of 100%, and a positive predictive value of 1,8% |
| Zhang et al. [47] |
2022 | 57 | ARVC/NDLVC | Transcriptome profiles from human hearts | SVM, naive Bayes, DT, K-NN, gradient-boosting machine, XGboost, and RF | R packages version 4.1.3 (R Foundation for Statistical Computing, Vienna, Austria) |
Inherently interpretable model Post-modeling explainability |
External validation | Public datasets (http://www.ncbi.nlm.nih.gov/geo; https://www.mdpi.com/2073-4425/11/12/1430/s1) accessed on 30 December 2023 No algorithm code available |
Random forest achieved the best performance, with an AUC of 0.86 in discriminating between arrhythmogenic cardiomyopathy and dilated cardiomyopathy. |
| Bleijendaal et al. [48] |
2020 | 310 | ARVC/NDLVC | ECG data | CNN, long short-term memory, K-NN, LR, multilayer perceptron, RF, SVM, XGboost |
Python version 3.9 (R Foundation for Statistical Computing, Vienna, Austria) |
Inherently interpretable model Post-modeling explainability |
External validation | No public datasets Algorithm code available (https://github.com/L-Ramos/CardiologyAI.) accessed on 30 December 2023 |
The proposed ML and DL models outperformed expert cardiologists in terms of accuracy and sensitivity. |
| Papageorgiou et al. [49] |
2022 | 183 | ARVC/NDLVC | ECG data | CNN | R packages version 4.1.0 (R Foundation for Statistical Computing, Vienna, Austria) |
Post-modeling explainability | Internal validation | No public datasets No algorithm code available |
The CNN model achieved 99.98% accuracy, 99.96% specificity, and 99.98% sensitivity during the training phase and 98.6% accuracy, 98.25% specificity, and 98.9% sensitivity when tested. |
| Chao et al. [50] |
2023 | 381 | RCM | TTE data | CNN | Python version 4.1.1 (Python Software Foundation, Beaverton, Oregon) |
Post-modeling explainability | External validation | No public datasets No algorithm code available |
The DL model yielded an AUC of 0.97 in differentiating constrictive pericarditis vs. cardiac amyloidosis. |
| Sengupta et al. [51] |
2016 | 94 | RCM | Clinical parameters, conventional TTE, and speckle tracking TTE data | Associative memory classifier, RF, K-NN, and SVM | R packages version 3.3 (R Foundation for Statistical Computing, Vienna, Austria) |
Post-modeling explainability | Internal validation | No public datasets No algorithm code available |
The proposed ML approach achieved an accuracy of 93.7% and an AUC of 96.2%. |
| Taleie et al. [37] |
2023 | 91 | RCM | radiomic features | K-NN, LR, multi-layer perceptron, RF, SVM, and XGboost | R packages version 4.0 (R Foundation for Statistical Computing, Vienna, Austria) |
Inherently interpretable model Post-modeling explainability |
Internal validation | No public datasets No algorithm code available |
The ML model achieved the best performance, with an AUC of 0.73, accuracy of 0.73, specificity of 0.73, and sensibility of 0.73. |
| Asmarian et al. [52] |
2022 | 624 | RCM | Clinical and laboratory data | RF, gradient boost model, and LR. | R packages version 4.1.0 (R Foundation for Statistical Computing, Vienna, Austria) |
Inherently interpretable model Post-modeling explainability |
Internal validation | No public datasets No algorithm code available |
The ML model yielded an AUC of 0.68 in evaluating heart iron overload. |
| Eckstein et al. [53] |
2022 | 96 | RCM | CMR strain and function parameters | k-NN, SVM, and DT | Python Version 3.8.12 (Python Software Foundation, Beaverton, Oregon) |
Inherently interpretable model Post-modeling explainability |
Internal validation | No public datasets No algorithm code available |
The ML-based model achieved an accuracy rate of 90.9% and an AUC of 0.996. |
| Cau et al. [54] |
2022 | 43 | Takotsubo syndrome | CMR parameters, demographics data | RF, bagging of trees, adaptive boosting, and XGboost | R packages version 4.1.0 Python version 3.9 (Python Software Foundation, Beaverton, Oregon) |
Inherently interpretable model Post-modeling explainability |
Internal validation | No public datasets No algorithm code available |
The extremely randomized trees ML algorithm showcased a sensitivity of 92% (with a 95% confidence interval of 78–100), a specificity of 86% (95% CI 80–92), and an AUC of 0.94 (95% CI 0.90–0.99) in the diagnosis of Takotsubo syndrome. |
| Izquierdo et al. [36] |
2021 | 118 | Left ventricle non-compaction | Radiomics features | SVM, RF, LR | Python version 3.7.9 (Python Software Foundation, Beaverton, Oregon) |
Inherently interpretable model Post-modeling explainability |
Internal validation | No public datasets No algorithm code available |
The radiomics models for the automated diagnosis of left ventricle non-compaction achieved excellent diagnostic performance, with AUC values of 0.95. |