ML basics in cardiology

ML applications in cardiovascular medicine employ various algorithmic approaches, each with its advantages for specific diagnostic questions. Supervised learning algorithms, such as support vector machines, random forests, and neural networks, have demonstrated excellent performance in classification tasks related to pattern identification, particularly in detecting CAD [15].

Deep learning (DL) architectures such as convolutional neural networks (CNNs) have changed medical image analysis by automatically extracting differentiating features from complex imaging data without significant manual feature engineering. CNNs can simultaneously process large amounts of imaging data and identify subtle patterns and relationships that are beyond human perception [16].

An overview of supervised, reinforcement, and unsupervised learning across ECG, magnetic resonance imaging (MRI), and echocardiography (ECHO) is shown in Figure 1, which illustrates how multimodal inputs feed algorithmic pipelines to generate diagnostic outputs.

Display full size

Figure 1. Algorithmic approaches in cardiovascular diagnosis. Schematic overview showing how multimodal inputs—ECG, cardiac magnetic resonance (MRI), and transthoracic echocardiography (ECHO)—feed algorithmic pipelines to produce diagnostic outputs. The figure contrasts three learning paradigms: (1) supervised learning (labeled data → model training → predictions and model update with new data); (2) reinforcement learning (agent-environment loop using state, action, and reward to improve policy); and (3) unsupervised learning (clustering/structure discovery from unlabeled data). The end-point is a workstation rendering of a cardiac diagnosis. ECG: electrocardiogram; MRI: magnetic resonance imaging. Created in BioRender. Tomala, M. (2025) https://www.biorender.com/7ykr1qa.

AI in electrocardiography

AI-facilitated electrocardiographic interpretation exemplifies a modern achievement of effective identification in the cardiovascular sphere. Early studies are also exploring AI-enabled ECG models to identify patients with silent ischemia or obstructive CAD despite normal ECG findings; however, this approach is not yet validated for routine practice [17–19]. DL networks have learned patterns from millions of ECG recordings and can detect subtle changes potentially associated with CAD, in some cases achieving performance comparable to experienced clinicians [16].

Advances in CCTA

CCTA has evolved from an anatomical imaging modality to a diagnostic tool through algorithmic advancements. Modern AI-based CCTA analysis platforms can determine the presence, quantity, and type of coronary plaque with high accuracy and reproducibility [20, 21].

ML algorithms applied to CCTA data can identify high-risk plaque features, such as positive remodeling, low-attenuation plaques, spotty calcifications, and the napkin-ring sign—all of which are linked with increased risk of acute coronary syndromes. These systems exhibit high consistency compared to human interpretation, saving time on analysis [8].

The clinical workflow and system architecture for AI-enabled CCTA analysis are summarized in Figure 2.

Display full size

Figure 2. AI-enhanced CCTA analysis protocol. Panel A—Clinical workflow & AI processing pipeline. Image acquisition (premedication as required; heart-rate control; ECG-gated CCTA with contrast; standard technical parameters and quality control) is followed by AI processing (automated coronary segmentation; plaque and stenosis detection/quantification using convolutional neural networks). Clinical output includes CAD-RADS 2.0 grading (categories 0–5), stenosis/severity assessment, risk stratification (e.g., MACE prediction with calibrated probabilities and confidence intervals), and structured reporting; integration options include HL7/FHIR and accelerated reporting. Panel B—System integration architecture. CT scanner (DICOM) → AI processing service → validation/quality check → structured report (DICOM SR TID 1500) → PACS archive → EMR for the finalized clinical report. AI: artificial intelligence; CAD-RADS: Coronary Artery Disease Reporting and Data System; CCTA: coronary computed tomography angiography; EMR: electronic medical record; FHIR: Fast Healthcare Interoperability Resources; HL7: Health Level Seven; MACE: major adverse cardiac events; PACS: picture archiving and communication system; CT: computed tomography; DICOM: Digital Imaging and Communications in Medicine; ECG: electrocardiogram.

By integrating FFR derived from CT (FFR-CT) with ML algorithms, the diagnostic capabilities of CCTA have advanced further. AI-enhanced FFR-CT analysis can determine the hemodynamic significance of coronary stenosis with similar diagnostic accuracy to invasive pressure wire measures, but non-invasively [9]. For example, the commercially available HeartFlow FFR-CT platform has been shown in multicenter studies to improve CCTA’s diagnostic performance by correctly reclassifying non-significant lesions and matching invasive FFR measurements [9].

Advances in cardiac MRI

By incorporating AI enhancements, CMR imaging enables the comprehensive assessment of myocardial anatomy, function, and tissue characteristics on a broader scale than previously possible. More recent automated analysis programs can segment cardiac chambers with minimal user input, measure ejection fraction, and detect myocardial scarring (e.g., myocarditis) with accuracy and timeliness [11].

Moving outside the typical CMR anatomical analysis, new DL models applied to CMR perfusion imaging identify subtle perfusion abnormalities associated with CAD, with improved sensitivity compared to visual analysis. These algorithms examine dynamic sequences of perfusion, including quantitative perfusion mapping, and quantify regional perfusion defects with high diagnostic accuracy [10].

The AI-integrated pipeline for chamber segmentation, tissue characterization, and perfusion mapping is depicted in Figure 3, which summarizes the CMR preprocessing, DL analysis, and standardized reporting steps.

Display full size

Figure 3. Automated cardiovascular magnetic resonance (CMR) analysis workflow with AI integration. Stage 1—DICOM download of multi-sequence CMR (cine, LGE, T1/T2 mapping, perfusion, etc.). Stage 2—AI preprocessing (denoising, motion correction, intensity normalization, quality assurance). Stage 3—DL analysis for automated LV/RV chamber and myocardial segmentation, feature extraction, perfusion, and strain metrics. Stage 4—Clinical parameters: ejection fraction, ventricular volumes, mass, strain, and flow/valve quantification. Stage 5—Clinical report: automated, standardized, and structured. System performance/implementation notes: processing designed to complete within minutes per study, supports batch processing, exhibits high segmentation accuracy and reproducibility, is DICOM-compatible, and integrates with HL7/FHIR and electronic medical record (EMR) systems. DICOM: Digital Imaging and Communications in Medicine; DL: deep learning; LV/RV: left/right ventricle; LGE: late gadolinium enhancement; AI: artificial intelligence; HL7: Health Level Seven; FHIR: Fast Healthcare Interoperability Resources.

Key AI applications, typical performance ranges, and primary use cases are summarized in Table 1. Table 1 provides a modality-by-modality overview relevant to screening, anatomic, and physiologic assessment.

High-definition IVUS (HD-IVUS)

HD-IVUS represents a significant advancement in intracoronary visualization strategies, providing an unprecedented understanding of coronary vessel wall morphology and plaque features. The integration of AI algorithms with HD-IVUS has transformed our understanding of plaque analysis and procedural guidance [24].

State-of-the-art HD-IVUS systems with ML capabilities can automatically identify vessel boundaries, quantify plaque burden, and characterize plaques in terms of their components. These algorithms can differentiate fibrous, fibro-fatty, necrotic core, and dense calcium features, providing a more thorough plaque analysis to guide therapy decisions [25].

The end-to-end AI pipeline for HD-IVUS acquisition, analysis, and reporting is shown in Figure 4. Figure 4 illustrates automated lumen/media segmentation, plaque characterization, and decision-support outputs.

Display full size

Figure 4. HD-IVUS algorithm analysis platform. HD-IVUS: high-definition intravascular ultrasound.

End-to-end workflow for automated analysis of HD-IVUS, showing:

• 1.

  Data acquisition—HD-IVUS scanner with automatic pullback, ECG-gated acquisition, Digital Imaging and Communications in Medicine (DICOM) interface, and storage of raw/RF frames.

• 2.

  Pre-processing pipeline—noise filtering, speckle reduction, image enhancement, temporal alignment, artefact suppression, motion compensation, and quality assessment.

• 3.

  AI analysis engine DL models (e.g., CNN, U-Net, ResNet, Transformer) performing lumen/media segmentation, plaque characterization, disease/stenosis detection, quantitative metrics (e.g., minimum lumen area, plaque burden), and risk-prediction algorithms.

• 4.

  Clinical output—three-dimensional (3D) vessel reconstruction, automated stenosis report, decision-support summaries for risk stratification and treatment guidance, picture archiving and communication system (PACS)/electronic medical record (EMR) integration, and structured reporting.

• 5.

  Continuous learning feedback loop—expert validation/annotation, quality-metrics monitoring, dataset curation, model versioning, clinical-outcomes tracking, and automated model updates.

Key performance domains. Segmentation accuracy, processing speed (per pullback/real-time capability), diagnostic performance [sensitivity/specificity/negative predictive value (NPV)/positive predictive value (PPV) by task], and clinical agreement with experts.

Platform infrastructure (implementation example): cloud-native deployment with containerization/orchestration for scalability, RESTful APIs with Health Level Seven (HL7)/Fast Healthcare Interoperability Resources (FHIR) interoperability, and DICOM connectivity to PACS systems.

The AVVIGO+ Automated Lesional Assessment platform is an example of the integration of AI with HD-IVUS imaging. This platform provides automated measurement of luminal area, vessel area, and plaque burden in real-time, with 85% agreement for lumen assessments and 97% agreement for stent area measurements [22].

Advanced neural networks applied to HD-IVUS data can predict procedural outcomes related to optimal stent size, expansion characteristics, and potential complications. These predictive models can analyze the overall geometry of the vessel, the distribution of plaque, and the characteristics of the underlying tissue to recommend personalized procedural strategies for each patient [26].

Enhancements to OCT

OCT is currently the highest-resolution imaging modality for coronary vessels, providing detailed information on plaque microstructure and the characteristics of the vessel wall. The advent of AI to enhance the analysis of images derived from OCT has enabled OCT to become a powerful diagnostic and therapeutic guidance tool [12].

ML applications to OCT data have automated the identification and quantification of thin-cap fibroatheromas (TCFAs), lipid pools, macrophage infiltration, and other components of high-risk plaque with greater accuracy than expert human analysis. Furthermore, these ML algorithms are capable of analyzing terabytes of data across thousands of OCT cross-section images in seconds, leading to a robust assessment of the entire vessel [13].

As shown in Figure 5, the OCT AI engine automates lumen and wall segmentation, feature extraction, and high-risk plaque characterization. Figure 5 displays the processing steps and clinical outputs used for percutaneous coronary intervention (PCI) planning and follow-up.

Display full size

Figure 5. AI-enhanced optical coherence tomography (OCT) analysis platform: automated intracoronary image processing workflow. Input—OCT acquisition. High-resolution intracoronary pullbacks in a DICOM-compatible format. Processing—AI engine (deep learning CNN). (1) Preprocessing & enhancement; (2) automatic lumen/wall segmentation; (3) feature extraction; (4) plaque characterization; optional (5) risk-score calculators and automated measurements. Screenshot—user interface view showing automated lumen and wall segmentation overlays, plaque characterization outputs, and stent-planning suggestions (sizing and landing zones). Output—clinical report. Automated identification of lesions and high-risk features (e.g., TCFA, macrophage signal, calcium burden), quantification, stent-planning suggestions, and longitudinal follow-up visualization. Clinical applications. PCI optimization (device sizing, landing-zone selection, post-stent assessment) and risk assessment (event prediction, restenosis monitoring). System performance & benefits. Fast processing, high reproducibility/standardization, improved measurement quality, and comprehensive structured reports. CNN: convolutional neural network; DICOM: Digital Imaging and Communications in Medicine; PCI: percutaneous coronary intervention; TCFA: thin-cap fibroatheroma; AI: artificial intelligence.

The Ultreon OCT software is a clinical example of AI-enabled software for OCT-guided interventions. The Ultreon system uses ML algorithms that automate lumen segmentation, make stent sizing recommendations, and identify optimal landing zones, with the aim of improving clinical outcomes.

Sophisticated OCT AI algorithms have emerged that can predict stent under-expansion for calcified lesions based on pre-procedural imaging analysis, which enables the identification of the best methods to address calcifications and plan for procedural optimization strategies. The predictive models of stent expansion in calcified lesions analyze calcification geometry and patterns, plaque morphology, and vessel geometry to inform deployment decisions, demonstrating significant potential [27].

Integrated diagnostic platforms

Currently, clinical decision support systems (CDSS) are state-of-the-art solutions that integrate various diagnostic modalities and utilize advanced AI technology to provide a comprehensive evaluation of cardiovascular health. These CDSS include an analysis of electrocardiographic data, pathways, and scores based on biochemical markers, imaging characteristics, and clinical details, providing personalized risk evaluations and treatment recommendations for patients [31].

In such decision-support applications, AI recommendations were found to be feasible and concordant with expert interventional cardiologists’ plans, highlighting that AI can assist without compromising decision quality. However, the ultimate judgment remains with the physician. A side-by-side comparison of AI vs. standard practice is shown in Table 2. Table 2 details endpoints, throughput, and cost implications by modality.

Precision medicine applications

Precision medicine involves utilizing the capabilities of healthcare best practices to support individuals. AI-supported diagnostic systems will enable us to receive individually tailored medicine through accurate phenotyping and risk-stratification algorithms. Predictive and prescriptive modeling can identify subpopulations of patients with similar features or presentations; each may have an idiosyncratic pathophysiology or require patient-specific therapeutics. It is not unreasonable to expect that ML algorithms, when applied to integrated clinical data, will have the ability to predict treatment outcomes for individual patients based on a specific intervention, create a personalized list for drug selection, or identify those at the highest risk for an unwanted outcome. Personally tailored medicine is the future of cardiovascular medicine [38].

Regulatory status, validation evidence, and implementation notes for representative systems are summarized in Table 3. Table 3 lists platforms, methods, and deployment characteristics.

Progress in DL

Progress in DL architecture, particularly transformer networks and attention-based architectures, is being applied to cardiovascular applications. These types of models have the potential to learn and generate sequential data, identify temporal dependencies, and incorporate multiscale information of unprecedented complexities [41].

Generative adversarial networks (GANs) have been explored for generating synthetic data to support robust algorithms with limited data availability. This, along with other technologies, may help accelerate the algorithm development process and facilitate its application to other populations [42].

Federated learning applications

Federated learning methods enable the collaborative development of algorithms across multiple institutions while safeguarding the privacy of patient data. These methods support the development of robust diagnostic models using a variety of data sources while precluding the necessity for centralized data sharing, all while addressing privacy concerns and complying with all relevant regulations [43].

Real-time implementation

The advancement of edge computing solutions will allow for the deployment of AI algorithms in real-time within clinical contexts. These designs can provide instant feedback on diagnostic outcomes, procedural assistance, and risk evaluation, eliminating the need for cloud-based processing functions [44].

Emerging methods and their potential advantages are summarized in Table 4. The table outlines technical descriptions and development stages.

Assessment of diagnostic accuracy and clinical validation

Meta-analyses indicate that AI-enhanced diagnostic systems can achieve diagnostic accuracies comparable to those of expert physicians in specific tasks. For example, pooled results for detecting CAD across various imaging modalities show sensitivities of ~ 85–95% and specificities of ~ 80–90%, which is on par with expert reader performance [46, 47].

The PROTEUS trial is the first prospective, randomized, controlled trial examining AI-assisted stress ECHO. This multicenter trial demonstrated the non-inferiority of AI-enhanced interpretation compared to standard clinical practice for coronary angiography referral decisions [40]. Notably, the PROTEUS results showed no significant difference in appropriate angiography referrals between AI-assisted and standard interpretations in the overall cohort, confirming non-inferiority. Significantly, AI support improved decision-making consistency among less-experienced clinicians and in complex cases, suggesting that AI can help level the field in diagnostic accuracy across operators [40].

Clinical implementation issues and barriers

While there is evidence of technical efficacy, several barriers hinder the implementation of AI diagnostic systems in clinical practice. These include approval processes for healthcare systems, acceptance by physician groups, reimbursement considerations, and integration with existing clinical workflows [48].

Another significant hurdle is standardizing AI algorithms across vendor platforms and clinical contexts. Comprehensive establishment of standardized protocols and comprehensive validation frameworks is necessary to facilitate widespread clinical adoption [47]. Another barrier is the lack of interpretability of many AI models—so-called “black box” algorithms—which can hinder trust among clinicians and patients. Ensuring transparency and explainability in AI decisions is increasingly seen as crucial for acceptance in practice. Data privacy and security also pose challenges; large datasets are needed to train robust algorithms, raising concerns about patient consent and data protection, especially when integrating AI tools across different hospital systems.

Food and Drug Administration (FDA) approval pathways

The regulatory landscape for AI medical devices is evolving as various government agencies with differing perspectives attempt to develop specific pathways for approving ML-based diagnostic tools. The FDA’s Software as Medical Device framework outlines the recommendations for algorithm validation and approval processes. As of 2023, the FDA has cleared or approved numerous AI-based medical devices, including several in cardiology, but the regulatory framework is still evolving to address ongoing algorithm updates and real-world performance monitoring [49].

Recent approvals of AI-enhanced cardiovascular diagnostic tools, including automated ECG interpretation systems and imaging analysis systems, have laid the groundwork for regulatory considerations and evaluation of this technology [50].

Ethical implementation

Ethical questions surrounding the implementation of AI in cardiovascular medicine include algorithmic bias, transparency requirements, and physician liability. To ensure equitable healthcare delivery, training datasets must be diverse and representative. Moreover, mechanisms for algorithmic accountability should be in place—including providing at least partial explanations for AI decisions and continuous monitoring of performance across different patient populations—so that biases can be detected and corrected [51].

Cost-effectiveness analysis

Early economic analyses suggest that AI-assisted diagnostic systems could lower healthcare expenditures through more efficient diagnosis, avoidance of unnecessary procedures, and improved resource utilization. There are reports of cost savings from 20–40% in health systems that incorporate an AI system into their daily clinical workflows [52]. However, these projections assume optimal implementation; real-world cost savings will depend on integration costs, staff training, and the extent to which such systems are adopted.

The improved time to interpretation, better consistency of diagnoses, and reduced repeat examinations offer significant economic benefits for health systems adopting AI diagnostics, resulting in a decrease in non-value-added time across the clinical workflow [53].

Workflow improvement

AI-assisted systems now save seconds on image assessment for each imaging study, rather than minutes, which does offer an improvement to workflow. Still, the actual impact on overall clinical workflow depends on factors such as system interoperability and the need for manual review of AI outputs—gains in speed must be weighed against time for validation of AI findings by clinicians. Automated preliminary reads have previously opened up the ideals of prioritizing urgent cases, reducing turnaround times, and optimizing resources [54].

Clinical and economic outcomes related to AI-assisted diagnostics are summarized in Table 5. Table 5 presents diagnostic accuracy ranges, projected cost savings, and time benefits.

A summary of performance across key applications is shown in Figure 6, which displays pooled estimates of sensitivity, specificity, and accuracy across different modalities and tasks.

Display full size

Figure 6. Performance metrics and clinical validation of AI applications in cardiovascular diagnostics. Descriptive comparison of the study-reported diagnostic performance by modality and task. Bars display the median of study-level sensitivity, specificity, and—where available—accuracy within each category; parentheses indicate the number of contributing studies: CCTA plaque detection (N = 19), OCT plaque erosion (N = 14), ML FFR-CT (N = 10), coronary angiography CNN models (N = 21), and CCTA stenosis detection (N = 7). Due to heterogeneous thresholds and reference standards, no formal meta-analysis was performed; values are descriptive and not pooled estimates. Where only one bar appears for a method, the other metrics were not consistently reported in the source literature. Key findings. OCT-based plaque-erosion detection showed the highest sensitivity (~ 96%) with strong overall balance across metrics; CCTA plaque and stenosis detection demonstrated consistently high accuracy across studies; ML FFR-CT achieved moderate performance (~ 76–80%) with variability by lesion complexity and image quality; CNN-based analysis of coronary angiography yielded high sensitivity (~ 90%). Overall variability across tasks underscores the need for application-specific validation and ongoing algorithm refinement before broad clinical deployment. CCTA: coronary computed tomography angiography; CNN: convolutional neural network; FFR-CT: fractional flow reserve derived from CT; OCT: optical coherence tomography; AI: artificial intelligence; CT: computed tomography.

A comprehensive analysis of AI performance across 12 key cardiovascular applications reveals notable differences in diagnostic accuracy, sensitivity, and specificity, depending on the specific clinical task. The top-performing applications include ECG arrhythmia detection (with 97.8% accuracy), heart disease prediction (with 97.6% accuracy), and cardiac MRI segmentation (with 95.2% accuracy). More difficult applications, such as FFR estimation, achieve lower but still clinically relevant performance levels of 84%. Implementation barriers and practical solutions across regulatory, ethical, clinical, and economic areas are outlined in Table 6. Table 6 offers a brief checklist for deployment planning.