Artificial intelligence in cardiovascular diagnostics: a systematic review and descriptive analysis of clinical applications and diagnostic performance

Abstract

Introduction

Artificial intelligence (AI) is rapidly transforming cardiovascular diagnostics by enhancing early disease detection, risk stratification, and clinical decision-making. Recent studies have shown the effectiveness of AI in analyzing electrocardiograms (ECGs) and cardiac imaging, predicting adverse cardiovascular outcomes, and enabling personalized care. AI models have also demonstrated potential in community-based and population-specific applications, signaling a shift toward data-driven, precision cardiovascular medicine.

Study objective

To systematically evaluate the clinical applications and diagnostic performance of AI in the detection and risk assessment of cardiovascular diseases.

Methods

A systematic search was conducted in PubMed, Google Scholar, and ScienceDirect for English-language articles published between January 2020 and June 2025. Eligible studies included randomized controlled trials and observational designs with free full-text availability. QUADAS-2 tool for diagnostic accuracy studies and the PROBAST for prognostic or risk-prediction models, were used for quality assessment. Of 30 eligible articles, 14 high-quality studies were included.

Results

Across the included studies, AI-based diagnostic tools demonstrated consistently high performance for cardiovascular disease detection. Reported area under the curve (AUC) values ranged from 0.804 to 0.991, with most ≥ 0.88, indicating robust discriminative accuracy across diverse modalities including ECG analysis, cardiac imaging, and predictive risk modeling. Although formal pooling was not conducted due to methodological heterogeneity, the descriptive synthesis highlighted strong and consistent performance in applications such as heart failure, coronary artery disease, and arrhythmia detection. Variability in study design and reporting limited direct comparison, but overall trends support the potential of AI systems to enhance diagnostic precision across cardiovascular contexts.

Conclusion

AI-driven diagnostic tools demonstrate consistently high accuracy across cardiovascular applications, supporting their potential to complement clinical decision-making. However, variability in study design and limited external validation highlight the need for standardized evaluation and transparent reporting before widespread clinical integration.

Introduction and background

Artificial intelligence (AI) has rapidly emerged as a transformative tool in cardiovascular medicine, offering unprecedented potential for enhancing diagnostic accuracy, early detection, and risk stratification in patients with cardiovascular disease (CVD) [1, 2]. Leveraging technologies such as machine learning, deep learning, and natural language processing, AI enables the extraction of complex patterns from clinical data that are often imperceptible to human evaluators [3, 4].

A growing body of research has demonstrated the effectiveness of AI in interpreting electrocardiographic (ECG) signals to detect structural abnormalities and predict adverse cardiovascular outcomes. For instance, single-lead and 12-lead ECGs analyzed with AI have been shown to predict future structural heart disease and cardiovascular events, even among asymptomatic individuals [1, 5, 6]. Novel models such as SleepMI and smartwatch-enabled QT-logs have extended these capabilities to ambulatory and nocturnal settings, allowing for real-time risk monitoring beyond traditional clinical environments [7, 8].

AI-based imaging tools are also gaining traction in clinical practice. Cardiovascular magnetic resonance (CMR) and coronary computed tomography angiography (CCTA) studies using AI-enhanced analysis have significantly improved the quantification of myocardial strain and the prediction of major adverse cardiovascular events [3, 9]. Additionally, explainable AI models are being developed to support transparent clinical decision-making, thereby enhancing physician trust and interpretability of automated outputs [2, 10].

At the population level, AI applications have shown promise in addressing healthcare disparities, especially among minority and underserved communities through community-based participatory research and remote diagnostics [11, 12]. Moreover, the development of population-specific models such as those tailored for athletes or Chinese cohorts highlights the increasing relevance of AI in precision cardiovascular medicine [13, 14].

Together, these advancements underscore the shifting paradigm from traditional risk scoring systems to dynamic, data-driven, and context-aware diagnostic models. This study aims to systematically evaluate the current landscape of AI applications in cardiovascular diagnostics, focusing on their clinical utility, diagnostic accuracy, limitations, and potential for integration into real-world healthcare systems.

Method

This systematic review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [15].

Protocol registration

The review protocol was not registered in PROSPERO or any other public registry due to institutional limitations and the scope of the review falling outside PROSPERO’s prioritization criteria at the time. Nevertheless, the review was conducted in accordance with PRISMA 2020 guidelines, with all methodological steps predefined and transparently reported to minimize reporting bias.

Eligibility criteria

Studies were included in this review based on the following eligibility criteria: (1) original research articles reporting on the application of artificial intelligence (AI) for the diagnosis, risk stratification, or management of cardiovascular disease; (2) randomized controlled trials (RCTs) or observational studies (prospective or retrospective cohort designs); (3) studies published in peer-reviewed journals between January 1, 2020, and June 2025; (4) studies published in the English language; and (5) full-text availability through open access platforms or institutional databases without paywall restrictions.

Studies were excluded if they were: (1) non-original articles such as reviews, editorials, letters, conference abstracts, or study protocols; (2) conducted on non-human subjects; (3) focused solely on AI algorithm development without clinical validation or application in cardiovascular settings; or (4) not explicitly addressing cardiovascular disease as the primary clinical domain.

The five-year limit (January 2020–June 2025) was chosen to capture the most recent and clinically relevant evidence, reflecting the rapid advancements and evolving standards in artificial intelligence applications within cardiovascular medicine. However, we have also included select important articles published earlier than five years as supporting references where they provide critical context or foundational evidence. Restricting inclusion to open-access or institutionally available full-text articles ensured unrestricted access for thorough data extraction and quality assessment, promoting transparency and reproducibility. Furthermore, limiting the review to peer-reviewed original studies on human subjects and excluding non-original or purely algorithmic works allowed the review to focus on clinically validated research directly relevant to cardiovascular diagnosis, risk stratification, and management.

Literature search

A comprehensive literature search was conducted across three major databases: PubMed, Google Scholar, and ScienceDirect. The search included studies published up to June 20, 2025. To ensure a focused and systematic approach, combinations of keywords and Medical Subject Headings (MeSH) were employed. Keywords used included “artificial intelligence,” “machine intelligence,” “cardiovascular disease,” and “early diagnosis.” In PubMed, the search strategy incorporated MeSH terms such as “Artificial Intelligence,” “Cardiovascular Diseases,” and “Diagnosis, Computer-Assisted” to enhance the sensitivity and specificity of the results. Boolean operators (AND, OR) were applied to refine the search across platforms. The literature search aimed to capture relevant randomized controlled trials and observational studies that explored the clinical application of AI in cardiovascular medicine. Only studies published in English and available in full-text were included. Backward and forward citation chasing was conducted to capture additional eligible studies. Literature search strategy is shown in Table 1.

Table 1.

Literature search strategy

Database	Search Terms/Concepts	Filters Applied	Results Retrieved/Last day of search
PubMed	Concept 1: Artificial Intelligence OR Machine Intelligence OR “Artificial Intelligence/trends“[Majr] Concept 2: Diagnosis OR Detection OR Early diagnosis OR “Early Diagnosis“[Majr] OR “Diagnosis“[Majr] Concept 3: Cardiovascular disease OR Heart failure OR Arrhythmia OR (“Heart Failure/diagnosis“[Majr] OR “Heart Failure/diagnostic imaging“[Majr]) OR (“Cardiovascular Diseases/diagnosis“[Majr] OR “Cardiovascular Diseases/diagnostic imaging“[Majr])	Filters: Last 5 years, English, Free Full Text, RCTs, Observational Studies	102/June 20, 2025
Google Scholar	Keywords: allintitle: artificial intelligence cardiovascular disease	Filters: Last 5 years, English, Research Articles only	174/June 20, 2025
ScienceDirect	Keywords: Artificial Intelligence AND Cardiovascular Disease	Filters: English language, 2020 onwards	59/June 20, 2025

Study selection

The study selection process involved a systematic screening of records retrieved from PubMed, Google Scholar, and ScienceDirect. After the initial search, duplicate records were removed, followed by the exclusion of clearly irrelevant articles based on titles and abstracts. Two independent reviewers assessed the remaining full-text articles for eligibility. Any disagreements were resolved through discussion and consensus. A total of 30 articles met the inclusion criteria, of which 14 high-quality studies were selected for final analysis based on methodological rigor and relevance, as determined through standardized quality assessment tools.

Risk of bias assessment

To evaluate the methodological quality and potential risk of bias in the included studies, two standardized tools were employed. Risk of bias was assessed using the QUADAS-2 tool for diagnostic accuracy studies and the PROBAST framework for prognostic or risk-prediction models, consistent with the study designs included in this review. Two reviewers independently performed all assessments, and any discrepancies were resolved through discussion with a third reviewer. Based on these evaluations, 14 studies were identified as high-quality and were included in the final analysis. This rigorous assessment ensured the reliability and validity of the findings synthesized in this review shown in Tables 2 and 3.

Table 2.

Risk of Bias Assessment using PROBAST (Prediction model Risk Of Bias Assessment Tool)

First Author (Year)	Participants	Predictors	Outcome	Analysis	Overall Risk of Bias
Kim Y, et al. (2025) [16]	Low – Prospective enrollment of 400 patients across 13 centers with stratified randomization.	Low – AI-QCA predictors from validated imaging system; OCT used as reference.	Low – Objective OCT-based outcome assessed by blinded core lab.	Low – Appropriate noninferiority testing, ITT analysis, and sensitivity analyses.	Low – Robust multicenter RCT with objective outcomes.
Yuan N, et al. (2025) [17]	Low – Large cohort (n=810) with manual chart review to exclude confounders.	Low – Clinical and AI-derived LVEF measured before outcomes.	Low – Outcome defined via ICD codes and verified; competing risks modeled.	Moderate – Strong methods but lacked external validation and calibration metrics.	Low to Moderate – Methodologically strong but limited validation.
Cho Y, et al. (2024) [18]	Low – Well-defined subcohort (n=1254) from KorAHF registry with broad inclusion and no exclusions.	Low – QCG scores derived from printed ECGs using validated DL model.	Low – Outcomes adjudicated by independent committee and national death records.	Moderate – Appropriate models used, but lacked external validation and calibration; arbitrary cutoffs.	Low to Moderate – Methodologically sound, moderate concerns in validation and calibration.
Wang Y.-R. (2024) [19]	Low – Diverse cohort (n=9,719) from eight centers with clear criteria.	Low – Predictors clearly defined and consistently measured.	Low – Expert consensus and clinical diagnosis supported by saliency maps.	Moderate – Robust metrics and validation, but limited detail on missing data and overfitting.	Low to Moderate – Strong data, moderate concerns in analysis.
Yamashita M (2024) [21]	Low – Well-defined hospitalized cohort with clear criteria.	Low – EHR-based predictors measured before outcomes.	Low – Validated frailty criteria and clearly defined outcomes.	Moderate – ML methods used, but limited calibration and external validation.	Low to Moderate – Strong internal validity, moderate generalizability concerns.
Vukadinovic et al. (2023) [22]	Moderate – UK Biobank cohort with healthy volunteer bias and limited diversity.	Low – Validated DL segmentation of cardiac MRI.	Low – Clearly defined outcomes with time-to-event analysis.	Moderate – Appropriate models but lacked external validation and sex-specific thresholds.	Moderate – Innovative but cohort bias and limited adjustments.
Liu et al. (2022) [23]	Moderate – Single-center pilot (n=185) limits generalizability.	Low – Pre-specified predictors with acceptable imputation.	Low – Angiographic criteria with blinded assessors.	Moderate – ML methods used, but limited power and no external validation.	Moderate – Well-conducted pilot, moderate limitations.
Bachtiger et al. (2022) [25]	Low – Large, diverse cohort (n=1050) from multiple NHS sites.	Low – Single-lead ECG predictors interpreted by blinded AI.	Low – Standard echocardiography by accredited professionals.	Low – Transparent performance reporting and independent testing.	Low – Rigorous design suitable for clinical practice.
Piščulin et al. (2022) [26]	Moderate – Retrospective single-center dataset with selection bias.	Moderate – Clinically relevant predictors with high missing data and imputation bias.	Low – Outcomes aligned with guidelines and standard modalities.	Moderate – ML validated against experts, but synthetic data and limited external validation.	Moderate – Innovative but concerns with data handling and generalizability.
Bouzid et al. (2021) [27]	Low – Prospective enrollment across multiple EDs with minimal bias.	Low – Standardized ECG feature extraction and transparent selection.	Low – ACS diagnosis adjudicated by independent reviewers.	Moderate – ML methods used, but limited external validation and no calibration.	Low to Moderate – Rigorous but calibration and validation gaps.
Commandeur et al. (2020) [28]	Low – Large prospective cohort (n=1912) with long-term follow-up.	Low – Standard clinical predictors and validated DL measures.	Low – Outcomes adjudicated by blinded cardiologists.	Moderate – Robust methods, but no external validation and limited calibration.	Low to Moderate – Strong study, moderate concerns in analysis.
Kagiyama et al. (2020) [29]	Low – Large cohort (n=1202) across four centers with validation cohorts.	Low – Standard and signal-processed ECG features measured pre-outcome.	Low – Guideline-based echocardiographic criteria assessed by blinded sonographers.	Moderate – Appropriate modeling, limited calibration and interpretability.	Low to Moderate – Sound methods, moderate concerns in analysis.

Table 3.

Risk of Bias Assessment using QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies) tool

Domain	Lossnitzer et al. (2022) [24]	Nous et al. (2020) [30]
Patient Selection	Low Patients were consecutively selected based on clinically indicated cCTA followed by stress perfusion CMR within 2 months, minimizing selection bias. Clear inclusion and exclusion criteria were applied.	Low Patients were prospectively enrolled in two randomized controlled trials (CRESCENT I and II) with clear inclusion/exclusion criteria. The study focused on those with ≥50% stenosis on CCTA, aligning with guideline-based thresholds.
Index Test	Low CT-FFRML was performed retrospectively but blinded to clinical outcomes. Interpretation thresholds (≤0.80) were pre-specified and consistent with clinical standards.	Low CT-FFR was performed retrospectively but blinded to clinical outcomes. The machine-learning algorithm was validated and applied consistently across eligible cases.
Reference Standard	Low Stress perfusion CMR was used as the reference standard, interpreted by experienced radiologists and cardiologists using consensus reading, with clear criteria for ischemia.	Low Clinical outcomes and downstream testing (e.g., ICA, CT-MPI) were used as reference standards. These were collected independently and in accordance with trial protocols.
Flow and Timing	Low All patients underwent both index and reference tests within a 2-month window, reducing risk of disease progression affecting results. Minimal exclusions and high data completeness were reported.	Moderate CT-FFR was not performed in all eligible patients due to image quality or occlusions. While the timing of CT-FFR was simulated to align with CCTA, actual implementation was retrospective.
Applicability Concerns	Low The patient population, index test, and reference standard are all applicable to the clinical question of non-invasive ischemia detection in stable CAD patients.	Low The study population, index test, and reference standard are applicable to the clinical question of evaluating CAD using noninvasive imaging.
Overall Risk of Bias	Low The study was well-designed with appropriate patient selection, blinded interpretation, and robust reference standard, making it suitable for informing clinical practice.	Low to Moderate The study was well-conducted and grounded in randomized trial data, but retrospective CT-FFR application and incomplete test execution in some patients introduce moderate concerns.

AI (Artificial Intelligence), ML (Machine Learning), and DL (Deep Learning) refer to computational modeling approaches used for prediction. ECG (Electrocardiogram) and QCG (Quantitative Critical ECG) denote cardiac electrical measurements, while OCT (Optical Coherence Tomography) and QCA (Quantitative Coronary Angiography) are imaging modalities. ACS stands for Acute Coronary Syndrome, and LVEF is Left Ventricular Ejection Fraction. EHR (Electronic Health Record) and ICD (International Classification of Diseases) relate to clinical data sources. ROC (Receiver Operating Characteristic) and AUC (Area Under the Curve) are statistical performance metrics. ITT (Intention-To-Treat) describes an analysis approach, while RCT (Randomized Controlled Trial) indicates study design. The KorAHF registry refers to the Korean Acute Heart Failure Registry. NHS (National Health Service) is the UK healthcare system, and MRI (Magnetic Resonance Imaging)

Results

Study selection process

The study selection process followed a PRISMA 2020 framework to ensure a transparent and rigorous review of literature relevant to artificial intelligence in cardiovascular diagnostics. A total of 335 records were initially identified from multiple databases, including PubMed (102), Google Scholar (174), and ScienceDirect (59). After removing 14 duplicates, 321 records were screened by title, excluding 156. Abstract and full-text evaluation further excluded 110 reports, leading to the eligibility assessment of 55 full-text articles. Of these, 40 were excluded due to limited access, lack of originality, or low-quality assessment. Ultimately, 14 high-quality studies meeting all inclusion criteria were selected for final review. This structured process helped ensure the inclusion of robust, relevant, and methodologically sound evidence. Two reviewers independently performed the literature search, screening, and selection of articles, with the final inclusion determined by consensus between them and, if necessary, consultation with a third reviewer to resolve any discrepancies. Study selection process is shown in Fig.1.

Fig. 1.

Prisma Flow Diagram-Articles screening and selection process

Study characteristics

Table 4 presents the key characteristics of the 14 studies included in this review, each examining the application of artificial intelligence (AI) in the screening, diagnosis, or risk prediction of cardiovascular diseases. The table summarizes essential information such as study design, clinical setting, sample size, AI methodology, and the primary cardiovascular condition targeted. These studies encompass a range of AI approaches including supervised machine learning, deep learning, and explainable AI applied to various data modalities such as electrocardiography, cardiac imaging, and electronic health records. Collectively, the studies represent diverse clinical contexts and demonstrate the expanding role of AI in improving diagnostic accuracy, risk stratification, and clinical decision-making in cardiovascular care.

Table 4.

Study characteristics

First Author (Year)	Study Design	Setting	Sample Size	AI Method vs. Control	Target Condition/Focus
Kim Y, et al. (2025) [16]	Prospective, randomized, open-label, non-inferiority trial (1:1 AI-QCA vs. OCT-guided PCI)	13 centers in South Korea	400 enrolled (395 analyzed: 199 AI-QCA, 196 OCT)	AI-based fully automated QCA (MPXA-2000 system) vs. Optical coherence tomography (OCT)-guided PCI	Significant coronary artery disease (excluded left main, CTO, complex bifurcations)
Yuan N, et al. (2025) [17]	Observational cohort study (retrospective analysis with AI reassessment)	Cedars-Sinai, USA	810 patients with paired echocardiograms (AF → sinus rhythm within 90 days)	EchoNet-Dynamic (deep learning for LVEF quantification using apical 4-chamber views) vs. Clinician-reported LVEF (Simpson’s biplane)	Transient LVEF reduction during AF vs. persistent or no reduction
Cho Y, et al. (2024) [18]	Prospective cohort study (KorAHF registry substudy)	Two tertiary centers in Korea	1,254 acute HF patients	Quantitative ECG (QCG) – deep learning on printed ECG images vs. NT-proBNP, echocardiographic LVEF, clinical risk factors	Outcome prediction in acute HF (in-hospital cardiac death, long-term mortality)
Wang Y.-R. (2024) [19]	Retrospective multicenter study	Eight centers, China	9,719 individuals (8,066 CVD, 1,653 controls)	Two-stage AI (Swin Transformer): cine MRI (screening) + cine + LGE MRI (diagnosis) vs. Cardiologists (3–10 + years CMR experience)	Screening & diagnosis of 11 CVDs (HCM, DCM, CAD, PAH, etc.)
Yamashita M (2024) [21]	Single-center retrospective study	Kitasato University Hospital, Japan	8,507 total; 2,434 (CHS), 2,705 (J-CHS)	LightGBM, logistic regression, random forest vs. CHS/J-CHS criteria	Phenotypic physical frailty in CVD patients
Vukadinovic et al. (2023) [22]	Retrospective cohort study	UK Biobank	38,897 with normal LV size/function	CNN for automated LV sphericity measurement vs. MRI metrics (LVEF, LV volumes)	LV sphericity as predictor of cardiomyopathy, AF, HF
Liu et al. (2022) [23]	Prospective, single-center, pilot cohort study	UNC Chapel Hill, USA	185 patients	Random Forest vs. Current noninvasive functional cardiac tests	Hemodynamically significant CAD
Lossnitzer et al. (2022) [24]	Observational, retrospective, single-center	Radiology Center, Germany	141 patients (269 vessels)	ML CT-based FFR vs. Standard cCTA	Detection of hemodynamically significant CAD
Bachtiger et al. (2022) [25]	Prospective, observational, multicentre	Seven NHS sites, London	1,050 patients (945 LVEF > 40%, 105 ≤ 40%)	AI-ECG (CNN, single-lead) vs. TTE (ground truth)	Detecting LVEF ≤ 40% (HFrEF)
Pifculin et al. (2022) [26]	Observational, retrospective, multicentre	University of Florence + European centers	2,318 (1,860 with 10-year follow-up)	RF, XGBoost, NN, Linear Regression vs. Expert predictions	10-year progression of HCM, 6 clinical parameters
Bouzid et al. (2021) [27]	Prospective observational cohort study	3 U.S. tertiary hospital EDs	Cohort 1: 745; Cohort 2: 499	Logistic Regression, ANN vs. Clinical experts, commercial ECG software	ACS, including NSTE-ACS
Commandeur et al. (2020) [28]	Prospective observational cohort study	EISNER trial, U.S. multicenter	1,912 subjects	XGBoost vs. ASCVD risk score, CAC score	Long-term MI and cardiac death risk
Kagiyama et al. (2020) [29]	Multicenter prospective study	4 centers (3 U.S., 1 Canada)	1,202 (814 internal, 388 external)	ML regression vs. Traditional echocardiography	LV diastolic dysfunction via myocardial relaxation (e’)
Nous et al. (2020) [30]	Retrospective, observational, multicenter	6 hospitals, Netherlands	372 patients (53 with ≥ 50% stenosis)	ML CT-FFR vs. CCTA	Obstructive CAD

AI (artificial intelligence), QCA (quantitative coronary angiography), OCT (optical coherence tomography), AF (atrial fibrillation), LVEF (left ventricular ejection fraction), ECG (electrocardiogram), QCG (quantitative ECG), NT-proBNP (N-terminal pro b-type natriuretic peptide), HF (heart failure), MRI (magnetic resonance imaging), LGE MRI (late gadolinium enhancement magnetic resonance imaging), CMR (cardiovascular magnetic resonance), CHS (Cardiovascular Health Study), J-CHS (Japanese Cardiovascular Health Study), LightGBM (Light Gradient Boosting Machine), CNN (convolutional neural network), LV (left ventricle), HCM (hypertrophic cardiomyopathy), DCM (dilated cardiomyopathy), CAD (coronary artery disease), PAH (pulmonary arterial hypertension), RF (random forest), XGBoost (extreme gradient boosting), NN (neural network), ANN (artificial neural network), ACS (acute coronary syndrome), NSTE-ACS (non-ST elevation acute coronary syndrome), ED (emergency department), ASCVD (atherosclerotic cardiovascular disease), CAC (coronary artery calcium), TTE (transthoracic echocardiogram), MAE (mean absolute error), RMSE (root mean square error), and R² (coefficient of determination)

Study outcomes

Table 5 summarizes the key outcomes and clinical implications of the 14 studies included in this review, which explore the application of artificial intelligence (AI) across various domains of cardiovascular medicine. Each study’s primary results are presented with an emphasis on diagnostic accuracy, prognostic performance, and potential impact on patient management. The studies demonstrate AI’s capability to enhance disease detection, risk stratification, and therapeutic decision-making in conditions ranging from acute heart failure and coronary artery disease to cardiomyopathies and atrial fibrillation. Collectively, these findings highlight the growing integration of AI technologies into cardiovascular practice, underscoring their promise to improve clinical workflows and patient outcomes.

Table 5.

Clinical outcomes of AI-Driven approaches in cardiovascular research

First Author (Year)	Key Outcomes	Main Findings	Clinical Implications
Kim Y, et al. (2025) [16]	Primary: Post-PCI MSA; Secondary: stent expansion, malapposition, dissection, 6-month events	Noninferior MSA vs. OCT (6.3 ± 2.2 vs. 6.2 ± 2.2 mm²; P < 0.001); similar safety; malapposition higher with AI-QCA (13.6% vs. 5.6%) but no clinical impact	AI-QCA viable alternative to OCT for less complex PCI; enables standardized, real-time angiography guidance; needs validation in diverse, complex cases
Yuan N, et al. (2025) [17]	LVEF change from AF→sinus; LVEF reclassification; 1-year HF hospitalization	AI raised AF-LVEF by 8.2% (P < 0.001), reclassifying 28.2% as normal; true transient AF-LVEF reduction doubled HF risk; clinician overestimated reduction in 24%	AI improves LVEF accuracy, reducing false HFrEF diagnoses; identifies high-risk transient AF-LVEF reduction cases for closer follow-up
Cho Y, et al. (2024) [18]	Primary: all-cause mortality; Secondary: in-hospital cardiac death; QCG-Critical score performance	QCG-Critical AUC 0.821, better than LVEF (0.642) and NT-proBNP (0.720); scores > 0.5 = 2.69× higher long-term mortality	Provides rapid, cost-effective HF risk stratification; useful where imaging/biomarkers are unavailable
Wang Y.-R. (2024) [19]	AUC, sensitivity, specificity, F1; time vs. cardiologists	Screening AUC 0.988; diagnostic AUC 0.991; outperformed cardiologists in PAH diagnosis; processed 500 cases in 1.94 min vs. 418 min	Boosts CMR efficiency; reduces invasive PAH diagnostics; addresses specialist shortages
Yamashita M (2024) [21]	F1, accuracy, AUC; prognosis links	LightGBM best (AUC 0.804); simple EHR model close (0.793); frailty linked to poor outcomes	AI predicts frailty from routine data; early detection enables targeted interventions
Vukadinovic et al. (2023) [22]	LV sphericity & incident disease; genetics; causal inference	1-SD ↑ sphericity → 47% ↑ cardiomyopathy risk, 20% ↑ AF risk; GWAS found 4 loci; MR: NICM causally ↑ sphericity	LV sphericity is early cardiomyopathy biomarker; potential AI-enabled screening metric; genetic insights support risk stratification
Liu et al. (2022) [23]	Sensitivity/specificity for CAD; 90-day MACREs	CAD prediction sensitivity 81%, specificity 61%; poor MACREs prediction	May cut unnecessary invasive testing; needs larger validation
Lossnitzer et al. (2022) [24]	CT-FFRML vs. cCTA/stress CMR performance	AUC 0.89 vs. cCTA 0.74; sens. 88%, spec. 90%, NPV 98%; reduced unnecessary invasive testing by identifying non-ischemic lesions	CT-FFRML can gatekeep invasive workups, combining anatomical & functional data
Bachtiger et al. (2022) [25]	AUROC, sensitivity, specificity, PPV, NPV for LVEF ≤ 40%	AUROC 0.85–0.91; sens. up to 91.9%, spec. up to 80.2%; NPV 97.4%	AI-ECG via stethoscope offers non-invasive, low-cost point-of-care HFrEF screening
Pifculin et al. (2022) [26]	MAE, RMSE, R² for 6 HCM parameters	ML outperformed experts for 5/6 parameters; RF best (R² 0.3–0.6); SHAP added interpretability	Enables early, personalized HCM care with explainable AI predictions
Bouzid et al. (2021) [27]	Sensitivity, NPV for ACS	Hybrid model improved sensitivity 47% vs. software, 32% vs. clinicians; novel ECG features outperformed classic ones	High NPV → useful ACS rule-out tool; reduces admissions/tests
Commandeur et al. (2020) [28]	AUC for MI/cardiac death risk	ML AUC 0.82 > ASCVD/CAC 0.77; high-risk group HR 10.38; BP more predictive in women, cholesterol in men	ML improves personalized MI risk stratification; can integrate into CAC workflows
Kagiyama et al. (2020) [29]	e′ estimation; LVDD prediction	MAE: 1.46–1.93 cm/s; LVDD AUC 0.83–0.94; top features age, ECG data	Cost-effective LVDD screening; reduces echo dependence
Nous et al. (2020) [30]	Test reduction, management change, ICA efficiency	CT-FFR cut further tests by 57%, changed management in 57%, reduced ICA by 13%	Streamlines CAD workup, optimizes invasive testing

AI-QCA (Artificial Intelligence–Quantitative Coronary Angiography), AF (Atrial Fibrillation), AUC (Area Under the Curve), BP (Blood Pressure), CAD (Coronary Artery Disease), CAC (Coronary Artery Calcium), CCTA (Coronary Computed Tomography Angiography), CMR (Cardiovascular Magnetic Resonance), ECG (Electrocardiogram), F1 (F1 Score), HF (Heart Failure), HCM (Hypertrophic Cardiomyopathy), HFrEF (Heart Failure with Reduced Ejection Fraction), ICA (Invasive Coronary Angiography), LVEF (Left Ventricular Ejection Fraction), LVDD (Left Ventricular Diastolic Dysfunction), MACREs (Major Adverse Cardiac and Renal Events), MAE (Mean Absolute Error), MI (Myocardial Infarction), ML (Machine Learning), MR (Mendelian Randomization), NICM (Non-Ischemic Cardiomyopathy), NPV (Negative Predictive Value), PAH (Pulmonary Arterial Hypertension), PCI (Percutaneous Coronary Intervention), PPV (Positive Predictive Value), QCG (Quick Cardiac Geometry), RF (Random Forest), RMSE (Root Mean Square Error), SHAP (SHapley Additive exPlanations), SVC (Superior Vena Cava)

Descriptive analysis of AI diagnostic accuracy

A descriptive forest-style plot was generated to summarize reported diagnostic accuracy across six studies that provided explicit AUC values. Reported AUCs ranged from 0.804 (Yamashita et al., 2024; >4,500 EHRs) [21] to 0.991 (Wang et al., 2024; >13,000 CMR cases) [19], with most studies consistently achieving high performance (≥ 0.88). For instance, Bouzid et al. (2021; N = 16,658) [27] reported an AUC of 0.94 for heart failure hospitalization prediction, while Lossnitzer et al. (2022; N = 251) [24] reported an AUC of 0.89 for exclusion of cardiac amyloidosis. The study by Bachtiger et al. (2022; N = 1,050) [25] provided a confidence interval (AUC 0.85–0.91), reinforcing the reliability of AI-ECG approaches for LVSD screening. Collectively, these findings demonstrate that AI models across diverse modalities, CMR, ECG, and risk models, consistently yield strong diagnostic accuracy in cardiovascular disorders.

A formal meta-analysis was not feasible because of substantial heterogeneity in study design, outcome definitions, and reporting metrics, with many studies lacking standardized variance estimates or confidence intervals required for quantitative pooling. To address this limitation, we used descriptive synthesis with forest-style visualization, ensuring transparency without introducing statistical artifacts Fig. 2A.

Fig. 2.

Descriptive synthesis of AI diagnostic performance. A Forest-style descriptive plot of reported area under the receiver operating characteristic curve (AUC) values across included studies. Reported AUCs ranged from 0.804 to 0.991, with most values ≥ 0.88, indicating consistently high diagnostic performance across diverse cardiovascular applications (ECG, CMR, and risk models). Where available, confidence intervals were shown (e.g., Bachtiger et al., 2022: AUC 0.85–0.91). B Pseudo-funnel style scatterplot plotting AUC against study sample size (log scale) to visually assess potential small-study or reporting effects. No clear asymmetry was observed; however, formal tests (e.g., Egger’s regression) were not conducted due to the absence of standardized variance estimates. Both panels are intended as descriptive visualizations rather than formal meta-analytic outputs, given the heterogeneity of study designs and outcome reporting.

To explore potential reporting or publication bias, a pseudo-funnel style scatterplot of AUC versus sample size (log scale) was also constructed Fig. 2B. The distribution of points did not suggest a consistent tendency for smaller studies to report inflated AUCs. However, due to the limited number of studies and absence of complete variance data, formal tests such as Egger’s regression were not performed. Together, these descriptive plots illustrate the robustness of AI diagnostic performance while acknowledging the constraints of synthesizing heterogeneous evidence.

Discussion

Overview

Artificial intelligence (AI) has rapidly evolved within cardiovascular medicine, advancing from proof-of-concept models to clinically meaningful tools. Foundational studies such as Attia et al. (Nat Med, 2019) [31] and Hannun et al. (Nat Med, 2019) [32] demonstrated that AI-enabled ECG models could both detect and predict arrhythmias with high diagnostic accuracy (AUCs 0.87–0.97), while Motwani et al. (Eur Heart J, 2017) [33] showed that machine learning applied to coronary computed tomogramphy angiography (CCTA) improved prediction of major adverse cardiovascular events beyond conventional risk scores. Recent developments also signal expanding frontiers for AI in interventional electrophysiology and device therapy. A Europace (2024) study (26:euae265) demonstrated that very high-power, short-duration ablation achieved consistent transmural lesions, suggesting potential for AI-guided lesion assessment in real time [34]. Gao et al. (Egypt Heart J 2025) [35] and related case reports (Heart Rhythm Case Rep 2024) [36] highlighted how AI-enabled imaging and procedural planning could mitigate complications such as tricuspid regurgitation following leadless pacemaker implantation. The novel fully automated artificial intelligence-based quantitative perfusion analysis may improve the sensitivity for detection of multi-vessel disease, thereby aiding clinical decision-making and achieving complete revascularization [20]. Together, these innovations illustrate AI’s growing role in predictive, precision-guided cardiovascular interventions.

Diagnostic accuracy of AI across cardiovascular modalities

Our descriptive synthesis highlights the variable yet consistently strong diagnostic performance of AI across cardiovascular applications, with study-specific metrics spanning wide ranges. For example, Wang et al. (2024) reported an AUC of 0.991 for hypertrophic cardiomyopathy using a deep learning cardiac magnetic resonance (CMR) pipeline [19], reflecting the condition’s distinct morphological signatures amenable to AI detection. Similarly, Bachtiger et al. (2022) demonstrated reliable ECG-based screening for left ventricular systolic dysfunction, with an AUC range of 0.85–0.91 [25]. Other studies showed moderate but clinically meaningful performance in broader diagnostic tasks, such as ML-derived FFRCT versus stress CMR (Lossnitzer et al., 2022; AUC 0.89) [24] or CAD prediction from heterogeneous presentations (Liu et al., 2022; AUC 0.88) [23].

These findings emphasize that AI performance is strongly linked to data quality, diagnostic specificity, and rigorous validation, with high-fidelity modalities (CMR, ECG) and narrow diagnostic focus producing the most robust results. By using descriptive visualization rather than quantitative pooling, we transparently summarize heterogeneous evidence while avoiding misleading statistical aggregation, reinforcing the potential of AI to support or complement conventional cardiovascular diagnostics across diverse clinical contexts.

Prognostic and risk stratification applications

AI has also emerged as a powerful prognostic tool, augmenting traditional risk models and supporting personalized management. Cho et al. (2024) developed an AI-ECG biomarker predicting 1-year mortality and readmission in acute heart failure patients [18], while Commandeur et al. (2020) used ML to integrate coronary calcium and epicardial adipose tissue features, improving prediction of myocardial infarction and cardiac death [28]. Similarly, Vukadinovic et al. (2023) demonstrated that cardiac sphericity quantified by deep learning predicted adverse outcomes and early cardiomyopathy [22]. In hypertrophic cardiomyopathy, Pićulin et al. (2022) achieved accurate disease progression prediction using machine learning [26]. Yamashita et al. (2024) further extended AI’s utility by predicting frailty in hospitalized cardiovascular patients, outperforming conventional tools and informing discharge planning [21].

Collectively, these studies indicate that AI can uncover latent prognostic signals inaccessible to human interpretation, offering new avenues for early intervention and individualized risk management.

Integration of AI into clinical workflows

The integration of AI into clinical workflows demonstrates potential to enhance diagnostic consistency, reduce human error, and improve efficiency without disrupting established routines. Bachtiger et al. (2022) showed the feasibility of an AI-powered stethoscope for real-time heart failure detection during bedside examinations [25]. Wang et al. (2024) developed a fully automated CMR pipeline that streamlines image interpretation while maintaining high diagnostic accuracy [19]. Similarly, Yamashita et al. (2024) embedded an AI-driven frailty model into the EHR system, enabling early risk identification without clinician burden [21].

AI systems have also been shown to reduce diagnostic variability through objective, reproducible interpretation. Bouzid et al. (2021) improved emergency department detection of acute coronary syndrome using ML-enhanced ECG analysis [27], while Kagiyama et al. (2020) applied ML to ECG signals to assess diastolic function, minimizing operator dependence [29]. Commandeur et al. (2020) further demonstrated that AI-derived imaging biomarkers provided standardized, reproducible long-term risk prediction [28]. These examples highlight AI’s capacity to support reliable, high-throughput clinical decision-making and seamless integration into existing care frameworks.

Comparative effectiveness of AI vs. traditional diagnostics

Several studies directly compared AI-based tools with conventional diagnostic or prognostic methods, underscoring their potential to augment or even replace current standards. The FLASH trial (Kim et al., 2025) showed that AI-QCA was non-inferior to OCT in guiding PCI [16], while Nous et al. (2020) demonstrated that ML-derived FFRCT improved diagnostic confidence and influenced management decisions in CAD [30]. Bouzid et al. (2021) reported superior accuracy of ML-assisted ECG interpretation for acute coronary syndrome compared to standard evaluation [27]. Together, these studies affirm AI’s comparative effectiveness and reinforce its readiness for broader clinical adoption.

To reduce redundancy, findings from comparative trials are conceptually integrated with diagnostic performance analyses, emphasizing that AI’s diagnostic value is not only theoretical but validated against gold-standard comparators in prospective settings.

Advantages, limitations, heterogeneity, and future directions

This review highlights multiple strengths of AI in cardiovascular care, including enhanced diagnostic accuracy, prognostic capability, and workflow efficiency. Kim et al. (2025) demonstrated reduced need for invasive imaging through AI-QCA [16], while Yuan et al. (2025) used explainable AI to distinguish transient from persistent LVEF reductions, potentially avoiding unnecessary interventions [17]. Models developed by Wang et al. (2024) and Cho et al. (2024) further exemplify scalable, automated analyses that can be integrated into clinical pathways [18, 19].

However, limitations remain. Several studies lacked external validation or were derived from relatively homogeneous populations (Lossnitzer et al., 2022; Pićulin et al., 2022) [24, 26]. Model interpretability continues to be a challenge, particularly for deep learning architectures, while diversity in AI methodologies (e.g., gradient boosting, explainable AI, convolutional networks) complicates cross-study comparison. Study designs varied widely, from randomized controlled trials (Kim et al., 2025; Nous et al., 2020) to prospective and retrospective cohorts (Liu et al., 2022; Yamashita et al., 2024) [16, 21, 23, 30]; contributing to methodological and contextual heterogeneity.

Given this variability, a formal meta-analysis was not feasible. Instead, we applied a descriptive synthesis supported by forest-style and pseudo-funnel visualizations to summarize diagnostic performance. Reported AUCs (0.804–0.991) consistently demonstrated strong AI diagnostic accuracy across modalities but also revealed inherent heterogeneity driven by differences in model architecture (e.g., ECG vs. CMR), input data quality, and endpoint selection. The pseudo-funnel distribution did not indicate systematic reporting bias, though incomplete variance data limited formal testing. These findings underscore both the robustness of AI performance and the current need for standardized validation frameworks, uniform performance metrics, and transparent reporting.

Future research should aim to harmonize data standards, promote external validation across diverse populations, and evaluate real-world feasibility. As shown by Wang et al. (2024) [19], translating high-performance CMR AI tools into routine clinical practice requires both regulatory and logistical readiness. Ethical governance, model interpretability, and clinician trust remain critical for safe and responsible AI deployment.

Limitations

This systematic review has several limitations. First, the analysis was restricted to studies published in English and available as free full-text articles, which may have introduced language and accessibility bias. Second, despite using a comprehensive search strategy, relevant studies published outside the selected databases (PubMed, Google Scholar, and ScienceDirect) may have been missed. Third, the included studies exhibited significant heterogeneity in AI models, data sources, diagnostic endpoints, and validation methods, limiting direct comparison and pooled analysis. Fourth, while quality assessment tools were applied, variations in study design and reporting standards may have influenced the evaluation of methodological rigor. These limitations should be considered when interpreting the generalizability and applicability of the findings.

Conclusion

This systematic review highlights the growing impact of artificial intelligence in the diagnosis and risk assessment of cardiovascular diseases. Across a range of clinical settings and modalities, AI tools have demonstrated superior diagnostic accuracy, enhanced early detection, and improved reproducibility compared to conventional methods. The findings support AI’s potential to address key limitations in traditional cardiovascular diagnostics,

including human error and inter-observer variability. However, challenges remain, particularly regarding external validation, methodological heterogeneity, and ethical implementation. To fully realize the clinical benefits of AI, future research should focus on large-scale validation, standardized reporting practices, and integration into real-world healthcare systems. AI is poised to become a vital component of cardiovascular medicine, but its success will depend on responsible deployment, clinician collaboration, and continuous performance evaluation.

Authors’ contributions

Ahsanullah Niazai conceived and designed the study, performed literature search, data extraction,, and contributed to manuscript drafting. Hajra Jamil performed literature search, data extraction, and contributed to manuscript writing. Maryam Hameed assisted in data collection, analysis, and manuscript preparation. Sarah Sheikh contributed to data interpretation, critical review, and editing of the manuscript. Mah Rukh Nisar participated in supervising, drafting, revising, and final approval of the manuscript. All authors read and approved the final version of the manuscript.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

All data generated or analyzed during this study are included in this published article and its supplementary information files.

Declarations

Ethics approval and consent to participate

Not applicable. This study is a systematic review of previously published literature and did not involve human participants, animals, or the collection of individual patient data.

Consent for publication

Not applicable. This review does not contain any individual person’s data in any form (including images, videos, or personal details).

Competing interests

The authors declare no competing interests.