Table 2.
Overview of AI use in Cardiology.
| Target | Type of Algorithm | Data Sample | Results | Study |
|---|---|---|---|---|
| Signal processing | ||||
| Detection of VF and VT (shockable rhythms) to improve shock advice algorithms in automated external defibrillators | Convolutional neural network as a feature extractor and boosting classifier | 1135 shockable segments and 5185 non-shockable segments from 57 records in public databases | Accuracy 99.3%, sensitivity 97.1%, specificity 99.4% | Nguyen et al., (2018) [57] |
| Automated detection of AF based on PPG and accelerometer recordings of smartwatches | Deep neural network with heuristic pre-training | Heart rate and step count data obtained using the Cardiogram mobile application on Apple Watches from 9759 Health eHeart Study participants | Sensitivity 98.0%, specificity 90.2%, C-statistic 0.97 | Tison et al., (2018) [58] |
| Binary classification ofcardiovascular abnormality using time–frequency features of cardio-mechanical signals, namely, SCG and GCG signals | Decision tree and SVM methods with features generated by a continuous wavelet transform | Experimental measurements from 12 patients with cardiovascular diseases and 12 healthy subjects | Accuracy > 94%, with the best performance of SVM applied to GCG features (99.5%) | Yang et al., (2018) [59] |
| Automated detection of AF based on Apple Watch Series 2 or 3 with KardioBand (AliveCor) | SmartRhythm 2.0, a convolutional neural network | Data of the heart rate, activity level, and ECGs from 7500 AliveCor users (training), and data from 24 patients with an insertable cardiac monitor and history of paroxysmal AF (validation) | Episode sensitivity 97.5%, duration sensitivity 97.7%, patient sensitivity 83.3% overall and 100% during time worn | Wasserlauf et al., (2019) [60] |
| Identify LV territory of regional wall motion abnormality on parasternal short-axis views | Convolutional neural networks (supervised) | In total, 400 patients (1200 short-axis echo videos) who had undergone a coronary angiography and echo | Area under the receiver operating characteristic curve for detection of regional wall motion abnormalities: 0.90–0.97 | Kusunose et al., (2019) [61] |
| Identification of asymptomatic LV dysfunction based on an ECG | Convolutional neural network using the Keras framework with a Tensorflow (Google) backend and Python | ECG–TTE pairs: 35,970 (training), 8989 (internal validation), 52,870 (testing) | Accuracy 85.7%, sensitivity 86.3%, specificity 85.7%, C-statistic 0.93 | Attia et al., (2019) [50] |
| Image processing | ||||
| Rapid and reproducible measurement of LV volumes, EF, and average biplane LS on ECG | Convolutional neural networks | Four- and two-chamber ECG views from 255 patients in sinus rhythm | Feasibility 98%, good agreements with the reference for automated EF and LS, with no variability | Knackstedt et al., (2015) [62] |
| Decreasing the computational demand of the FFR calculation by developing a ML-based model as an alternative to computational fluid dynamics approaches | Deep neural network | In total, 125 lesions in 87 patient-specific anatomic models generated from CT data using image segmentation | Excellent correlation (0.9994; p < 0.001) and no systematic bias in the Bland–Altman analysis: FFR 0.80 was predicted with sensitivity 81.6%, specificity 83.9%, accuracy 83.2% | Itu et al., (2016) [63] |
| Automated ECG interpretation, including view identification, segmentation of cardiac chambers across five commonviews, quantification of structures and function, and disease detection | Convolutional neural networks | In total, 14,035 echocardiograms spanning a 10-year period | Identification of views in >95%, median absolute deviation of 15–17% for structure and <10% for function, detection of hypertrophic cardiomyopathy, cardiac amyloidosis, and pulmonary disease with C-statistics of 0.93, 0.87, and 0.85, respectively | Zhang et al., (2018) [64] |
| Measurement of RV and LV volume and function in MRI images for a range of clinical indications and pathologies | Deep neural network | In total, 200 non-congenital clinical cardiac MRI examinations | Strong correlations for LV (>0.94) and RV (>0.92) volumes | Retson et al., (2020) [65] |
| Detection of subclinical AF | Convolutional neural networks | Training set of 454,789 images and testing on 130,801 images | AUC 0.90, sensitivity 82.3%, specificity 83.4%, accuracy 83.3% | Alzubaidi et al., (2021) [66] |
| Clinical risk stratification | ||||
| Identification of HF cases from both structured and unstructured EMRs | Random forest models | In total, 2,139,299 notes in the Maine Health Information Exchange EMR database from 1 July 2012 to 30 June 2014 | Positive predictive value of 91.4% | Wang et al., (2015) [67] |
| Development of CHIEF to automatically extract LV function mentions and values, congestive HF medications, and documented reasons for a patient not receiving these medications | Combination of rules, dictionaries, and ML methods | Various clinical notes from 1083 Veterans Health Administration patients | High recall and precision for HF medications and EF (>0.960), while only reaching fair recall and precision for reasons for not prescribing HF medications (<0.400) | Meystre et al., (2017) [68] |
| Risk prediction model of incident essential hypertension within the following year | Feature selection and generation of an ensemble of classification trees with the use of XGBoost | Data from individual patient electronic health records as part of the Health Information Exchange data set of Maine | C-statistics of 0.917 in the retrospective cohort and 0.870 in the prospective cohort | Ye et al., (2018) [69] |
| Predict survival following a routine echo using clinical and structured echo report data | Nonlinear random forest classifier (supervised) | In total, 171,519 patients (331,317 echo studies) using 90 clinical variables, LVEF, and 57 echo measurements. Labels were from clinical data and reported echo measurements | Area under the receiver operating characteristic curve: 1-year mortality, 0.85 5-year mortality, 0.89 |
Samad et al., (2019) [70] |
| Predict in-hospital mortality following an echo in patients admitted with heart disease using echo report data | Deep neural network (supervised) | In total, 25,776 in-patients admitted with heart disease based on ICD-10 codes. Labels were from clinical data and reported echo measurements | Area under the receiver operating characteristic curve: Overall, 0.90 Coronary heart disease subgroup, 0.96 Heart failure subgroup, 0.91 Area under the precision–recall curve, 0.28 |
Kwon et al., (2019) [71] |
| Prediction of CAD on CTA | Boosted ensemble algorithm | Clinical, CTA (CACS) in 13,054 subjects | AUC 0.881 | Lu et al., (2022) [72] |
AF, atrial fibrillation; CHIEF, Congestive Heart Failure Treatment Performance Measure Information Extraction Framework; CT, computed tomography; ECG, electrocardiography; EF, ejection fraction; EMR, electronic medical record; FFR, fractional flow reserve; GCG, gyrocardiography; HF, heart failure; LS, longitudinal strain; LV, left ventricular; ML, machine learning; MRI, magnetic resonance imaging; PPG, photoplethysmography; RV, right ventricular; SCG, seismocardiography; SVM, support vector machine; TTE, transthoracic echocardiography; VF, ventricular fibrillation; VT, ventricular tachycardia.