Table 1. List of studies conducted on AI-assisted diagnosis in cardiovascular diseases.

AUC: area under the curve; CAD: coronary artery disease' ANN: artificial neural network; SPECT: single-photon emission computed tomography; ROC: receiver operating characteristic; RCA: right coronary artery; LAD: left anterior descending artery; LCX: left circumflex artery; ICM: ischemic cardiomyopathy; DCM: dilated cardiomyopathy; cMRI: cardiac magnetic resonance imaging; BED: biomarker extraction and discovery; WEKA: Waikato Environment for Knowledge Analysis; BRL: Bayesian rule learning; sECG: subcutaneous ECG; TNF-α: tumor necrosis factor alpha; IL-2: interleukin-2; NT-proBNP: N-terminal pro b-type natriuretic peptide; AI: artificial intelligence; MESA: Multi-Ethnic Study of Atherosclerosis

Study	Study Design	Methodology	Outcomes
Ross Upton, 2022 [38]	Prospective multicenter   randomized crossover reader study	Evaluation of how availability of an AI classification might impact clinical interpretation of stress echocardiograms	Acceptable accuracy in identifying patients with severe CAD, heightened sensitivity in disease detection by 10%  resulting in an AUC of 0.93,  specificity of 92.7%, and a sensitivity of 84.4% enhance accuracy, inter-reader agreement, and  reader confidence
Ying Guo, 2023 [41]	Prospective randomized controlled  trial	The study included 818 patients undergoing coronary angiography, randomly divided into training (80%) and testing (20%) groups, with an additional 115 patients in the validation group. The study optimized a superior CAD diagnosis model using 59 echocardiographic features in a gradient-boosting classifier.	Characteristic AUC value of 0.852 in the test group and 0.834 in the validation group.  High sensitivity (0.952) and low specificity (0.691) effectiveness in detecting CAD   potential for increased false-positive
Hiroto Yoneyam, 2019 [42]	Prospective cohort study	The study included 59 patients diagnosed with stable CAD who had recently undergone both coronary angiography and myocardial perfusion SPECT imaging. The ability to identify culprit coronary arteries was evaluated for both experienced nuclear cardiologists and the ANN. This assessment was conducted using ROC curves and AUC analysis, allowing for a comparison of diagnostic accuracy between human experts and the AI system.	Diagnostic Accuracy: Observer A's accuracy with hybrid images: RCA: 83.6%,  LAD: 89.3%, LCX: 94.4%; Observer B's accuracy: RCA: 72.9%,  LAD: 84.2%, LCX: 89.3%; ANN's accuracy: RCA: 79.1%, LAD: 89.8%, LCX: 89.3%. Comparative Performance: the ANN demonstrated comparable diagnostic accuracy to experienced nuclear medicine physicians. Improvement with hybrid images: Significant enhancement in AUC for RCA region: Observer A: 0.715 to 0.835 (p = 0.0031), Observer B: 0.771 to 0.843 (p = 0.042).  Challenges: Identifying culprit coronary arteries from perfusion defects in the inferior wall without hybrid images was difficult due to individual variations in LCX and RCA perfusion areas.
Mei Zhou, 2023 [46]	Retrospective study	The study analyzed echocardiogram data from 399 patients (200 with DCM, 199 with ICM) who underwent angiography between 2016 and 2022 at a single hospital. An external validation group of 79 patients was included. Data were randomly split and analyzed using four machine-learning methods. Cross-validation was conducted within the primary cohort, and the external cohort tested model generalizability, enhancing the study's validity and potential clinical applicability.	XGBoost emerged as the best-performing method, achieving an AUC of 0.934, 72% sensitivity, 78% specificity, and 75% accuracy in the primary cohort. In external validation, it maintained robust performance with an AUC of 0.804, 64% sensitivity, 93% specificity, and 78% accuracy. The model demonstrated high discriminative ability, correctly identifying ICM with 72% sensitivity and DCM with 78% specificity.
Vanathi Gopalakrishnan, 2015 [47]	Retrospective study	The researchers developed and tested cMRI-BED, a novel informatics framework for biomarker extraction and discovery from complex pediatric cMRI data, applying it to 83 de-identified cases and controls to classify cardiomyopathy findings in children. The framework incorporates image processing, marker extraction, and predictive modeling tools, utilizing Bayesian rule learning for interpretable models and machine learning methods from the WEKA toolkit for performance assessment using accuracy and AUC measures	The BRL decision tree model achieved the best predictive performance with 80.72% accuracy and 79.6% AUC in 10-fold cross-validation. Notably, the model identified myocardial delayed enhancement (MDE) status as an important predictive variable, aligning with its known clinical significance in cardiomyopathy classification.
Fabio Quartieri, 2023 [48]	Retrospective study	This study aimed to evaluate the capability of an AI algorithm to expand ICM arrhythmia recognition beyond the standard four cardiac patterns. To achieve this, researchers conducted an exploratory retrospective analysis using sECG data.	AI can process sECG raw data coming from ICMs without previous training, extending the performance of these devices and saving cardiologists' time in reviewing cardiac rhythm pattern detection.
Caiwei Zhang, 2020 [49]	Retrospective study	The study analyzed 14 characteristics of heart disease patients in Cleveland and  Switzerland using various types of neural networks and classifiers to predict whether or not a patient has heart disease.	The logistic regression classifier performed better than other methods in predicting cardiovascular events.
Bharath Ambale-Venkatesh, 2017 [18]	Retrospective study	Used random survival forests, a machine learning method, to predict six different cardiovascular outcomes and compared its performance against traditional cardiovascular risk scores over 12 years. It included 6,814 participants (from the MESA) aged 45 to 84, with diverse ethnic backgrounds across the US, and focused on how early-stage disease progresses to cardiovascular events  in initially healthy people	Imaging, electrocardiography, and biomarkers were more predictive than traditional risk factors. Age was consistently the strongest predictor for overall mortality. Fasting glucose levels and carotid ultrasound measurements were key for predicting strokes. The coronary artery calcium score stood out for predicting coronary heart disease and other related cardiovascular issues. Measures of left ventricular function and cardiac troponin-T were crucial for predicting heart failure. Creatinine levels, age, and ankle-brachial index emerged as top predictors for atrial fibrillation. Biomarkers like TNF-α, IL-2 soluble receptors, and NT-proBNP were important across all outcomes  The random survival forests method outperformed traditional risk scores, improving prediction accuracy by reducing the Brier score by 10%–25%.