Artificial Intelligence for Cardiovascular Care in Action

Abstract

Artificial intelligence (AI) is catalyzing a new era in cardiovascular care. Innovations include the detection of latent signatures enabling early diagnosis, real-time nudges for optimized clinical decision-making, and digital twin simulations to personalize treatment. The integration of AI in health systems has the potential to improve patient outcomes and enhance the efficiency of care delivery across diverse populations. However, many challenges remain—including interoperability and data privacy, algorithmic fairness and sustainability, and system considerations such as workflow integration and policy. Here, we draw examples from state-of-the-art AI applications to examine the interplay between AI and learning health systems to augment continuous measurement and feedback, and implementation science to robustly evaluate the efficacy and safety of AI interventions deployed into clinical workflows. We share a vision for “AI in action”, transitioning cardiovascular health systems from sparse, reactive, and hospital-centric episodes to a multimodal, connected ecosystem continuously learning to improve patient outcomes.

Central Illustration

Highlights

• •AI-enabled technologies can act as a copilot for cardiovascular care, digitally enhancing perception, cognition, and action in a learning health system.
• •Implementation science principles can be applied to rigorously evaluate AI, establish guardrails against harmful learning, and operationalize tools for diffusion across cardiovascular care pathways.
• •We envision “AI in action” as the democratized implementation of trustworthy, adaptive AI within multimodal, connected ecosystems that continuously measure, learn, and improve cardiovascular outcomes.

implementing a (deep) learning health care system

Artificial intelligence (AI) systems have evolved to perform tasks once thought to require human cognition, including visual perception, understanding language, problem-solving through reasoning, and predicting outcomes from a range of actions. AI relies on pattern recognition algorithms that use vast data sets as a fundamental departure from older, rule-based engines. One form of AI, deep learning, particularly excels at identifying patterns in complex medical data.¹ AI applications in health care are propelled by the large and growing amount of digital, multimodal data—electronic health records (EHRs), the digitalization of diagnostic testing, and connected devices such as wearables—that train and power these models.² As such, AI seems a key and important tool in improving patient care and health system efficiency in cardiovascular care.3, 4, 5

However, a wide gap separates developments in health care AI from safe clinical adoption. For example, more than 40% of the AI-enabled devices cleared by the U.S. FDA (Food and Drug Administration) received regulatory authorization without clinical validation data,⁶ with recall rates exceeding 15%.⁷ In addition, rapid AI adoption globally has not followed implementation science best practices, which limits the broadest impact. More than 300 hospitals across China have installed on-site versions of the DeepSeek AI language model—integrated into diagnostics, decision support, and patient education—even as governance and regulatory approaches remain nascent.⁸ In the United Kingdom, 1 in 5 physicians reports using generative AI tools in clinical practice.⁹ Taken together, this depicts an urgent challenge, integrating best implementation practices to deploy AI safely and effectively for widespread clinical use in real-world settings.

A learning health system (LHS) is a concept that aims to improve health care by integrating scientific evidence directly into clinical practice, while simultaneously generating new knowledge from real-world care experiences.¹⁰ It is often described as a cyclical process where data are converted into knowledge, this knowledge is translated into practice, and changes in practice then generate more data, thus fostering continuous improvement.¹¹ Accomplishing the goals of an LHS requires technological innovation, aligned organizational incentives,¹² and a key focus on implementation science, which is a dynamic field dedicated to systematically studying methods to promote the integration of evidence-based practices, interventions, and policies into routine clinical settings.¹³ The fundamental purpose of implementation science is to bridge the gap between what is known from research and what is put into practice.¹⁴ This gap can be substantial; only 1 in 5 evidence-based interventions makes it to routine clinical practice, and the evidence generated from clinical research may take 17 years to be implemented.¹⁵ Applying LHS principles to AI implementation in cardiovascular care will involve prospective deployment, iterative evaluations of patient and process outcomes, and feedback loops that update models based on real-world performance and stakeholder experiences. This operationalizes implementation as a safe, effective, and scalable learning cycle. Several frameworks and comprehensive guidelines have been proposed for the implementation of AI in an LHS (Table 1).

Table 1.

Selected Theories, Models, and Frameworks Proposed to Bridge the Implementation Gap for AI in a Learning Health System

Framework	Summary of Guidelines
FUTURE-AI16	Fairness, universality, traceability, usability, robustness, and explainability. Defines 30 best practices addressing technical, clinical, and socioethical dimensions.
TRAIN-Europe17	Trustworthy and responsible AI network. Aimed at enhancing the quality of health care AI while promoting international collaboration and democratizing responsible AI. Proposes 5 areas to assess a hospital’s preparedness for responsible AI: AI model inventory, real-world data testing, bias assessment, governance, adoption and scalability.
NAM AICC18	National Academy of Medicine AI code of conduct. Aimed at harmonizing AI principles and mapping these to the National Academy of Medicine’s LHS shared commitments, with a focus on AI lifecycle, governance model, and stakeholder roles
ESC DCAI19	European Society of Cardiology’s digital cardiology and AI committee’s roadmap. Ensuring data quality and integrity, standardizing disease definitions and outcomes, creating a robust evaluation framework
CFIR 2.020	Consolidated framework for implementation research. Focuses on domains like implementation effectiveness, intervention characteristics, and settings.
RE-AIM21	Reach, efficacy, adoption, implementation, and maintenance. Guides the evaluation of process and health outcomes.
Diffusion of innovations22	Describes how a new practice spreads through a population, with key players being innovators, early adopters, early majority, late majority and laggards
TEHAI Framework23	Translational evaluation of health care AI. Emphasizes capability, utility, and adoption, with 15 subcomponents addressing technical validity, ethical safeguards, generalizability, safety, privacy, and real-world integration.
AI-TAM24	AI technology acceptance model. Extends the technology acceptance model with constructs for trust in AI, perceived AI output quality, and collaborative intention.

AI = artificial intelligence; LHS = learning health system.

In this review, we use state-of-the-art cardiovascular AI applications as case studies to illustrate the key themes of this special issue—AI for augmenting an LHS and implementation science for robust evaluation of AI interventions (Central Illustration). The ultimate vision is of “AI in action”: maximizing the utility of AI in an LHS with democratized access to widely implemented and trustworthy medical AI.

Central Illustration.

Implementing a (Deep) Learning Health System for Cardiovascular Care

AI as a copilot for cardiovascular care

AI tools can automate measurement and reporting across both structured and unstructured data, uncovering latent patterns that capture current disease trends and forecast future risk. Together, these capabilities enable high-fidelity clinical data abstraction, system-wide real-time quality measurement, and a shift from reactive to proactive care within an LHS (Figure 1).

Figure 1.

AI as a Copilot for Cardiovascular Care

AI augments the learning health system across four layers—measurement, reporting, detection, and prediction. For measurement and reporting tasks, NLP, LLMs, and RAG convert unstructured EHR text and imaging into structured abstractions, and image-to-text models can auto-generate diagnostics, streamlining quality measurement and documentation. For detection, CNNs infer latent disease signatures from routine signals. For prediction, feed-forward and transformer models—alone or fused with multimodal data such as EHR and CMR—forecast clinical trajectories and deliver individualized risk assessments. Collectively, these capabilities shift cardiovascular care from reactive episodes to proactive, system-wide decision support. CMR = cardiac magnetic resonance; ECG = electrocardiogram; EHR = electronic health record.

Autocompletion of measurement and reporting tasks

To evaluate care outcomes, health systems rely on process and outcome measures that span individual patients to hospital-level aggregate measures. Measurements rely on structured data (eg, patient billing codes for conditions, medication records, and laboratory values) and unstructured data (eg, text from clinical notes and medical images). Although structured data are easy to query, extracting key clinical information from unstructured data requires manual abstraction, with personnel and infrastructure costing millions of dollars annually.²⁵

AI systems can support these measurement tasks, aligning directly with the core objectives of an LHS, which aims to learn from daily experience and continuously improve care delivery.²⁶ Specifically, AI algorithms can automate data collection from disparate sources, enabling health systems to gain a comprehensive, accurate view of factors influencing health and disease, while reducing the burden on health care professionals.^10,11,18 In addition, AI algorithms can identify nonlinear relationships between these factors that traditional statistical methods might otherwise overlook^1,27,28 and facilitate better tracking of long-term outcomes and cross-institute benchmarking,^18,29 supporting an LHS’s goal of assessing and enhancing health quality.¹⁰

Advances in natural language processing (NLP) have enabled parsing of information that is traditionally captured in clinical notes. NLP further enhances the ability to not only extract words when they are mentioned but also provides broader inference from the context of the presented information. This is accomplished through neural networks with attention mechanisms to encode text into representations with meaning (semantic representations). These models are generally task-specific, such as identifying patients with aortic stenosis³⁰ or subtypes of heart failure (HF),31, 32, 33, 34 quantifying disease severity (eg, grading the functional status³⁵ or phenogroup³⁶ of patients with HF), or determining clinical endpoints (eg, adjudicating HF hospitalizations³⁷).

Newer implementations of language models include transformer-based large language models (LLMs), such as generative pre-trained transformer and Llama. These are general-purpose models trained on large and diverse corpora of text, building even more comprehensive semantic embeddings—numeric maps of how words and ideas relate. Using these maps, LLMs can generate representation-informed answers to tasks based on meaning and context, rather than simple keyword matches. LLMs have also been deployed to extract nonclinical data, such as social determinants of health, and to summarize clinical documentation.^38,39 Although LLMs have been developed using trillions of bytes of data across the internet, they require explicit adaptation for health system applications, including fine-tuning on medical text, incorporating their use in the context of the broader available health system data, and being grounded in medical literature to enable their accurate use. This limits their scalability due to the large computational and operational systems expertise needed to tune the model for each health care system and task.

One approach to this limitation is the use of retrieval-augmented generation (RAG) techniques. RAG is a query-based enhancement of LLMs that can enable medical adaptation.⁴⁰ RAG-LLMs store contextually-aware data embeddings that can be optimized for clinical text, allowing interoperable querying across database architectures. RAG-LLMs could be deployed across 2 different health system architectures to assess clinical quality measures such as stroke risk and bleeding risk in patients with atrial fibrillation (AF)⁴⁰ or identify reasons where therapeutic compliance is determined by patient preference or other subjective features. In addition to improved sensitivity and accuracy over structured data, the RAG-LLM also provided reasoning for its decisions. RAG-LLMs have also demonstrated scalability across an entire health system to accurately identify sepsis clinical quality care measures at the point of care.⁴¹

Beyond the reporting of quality-of-care assessments, other AI applications can further enhance the workflow of clinicians. Computer vision, that is, AI tools for imaging, can automate the inference of human-readable labels, such as the detection of severe aortic stenosis from echocardiography,⁴² or the identification of major coronary vessels and regions of stenosis on coronary angiographic images.⁴³ Furthermore, emerging innovations in image-to-text AI models can generate accurate, comprehensive reports from electrocardiograms (ECGs)⁴⁴ or echocardiograms,^45,46 streamlining diagnostic workflows. Together, these advances illustrate how AI can augment an LHS by alleviating reporting tasks and supporting health care workers.^47,48 The automation of reporting tasks using AI technologies enables clinicians to refocus on direct patient care, potentially enhancing professional well-being and supporting a culture of continuous learning within an LHS.¹⁰

Superhuman AI for the detection of latent disease signatures

AI, particularly via implementations of deep learning, excels at identifying subtle disease signatures within massive, complex, and high-dimensional datasets, often beyond the perceptual limits of human experts.^1,29 This “superhuman” capability to detect and predict disease status is poised to augment a health system as an engine for knowledge generation, enabling the extraction of meaningful insights from the vast stream of health data generated daily.^11,28 Furthermore, by accelerating the bidirectional evidence-to-practice learning cycle,⁴⁹ AI tools can potentially improve care delivery.²⁷

There are specific implementations of deep learning, such as convolutional neural networks, that enable automated pattern recognition in images, applying cascades of learnable convolutional filters and pooling layers to transform pixel data into hierarchical representations of features within images. The inference of complex disease signatures can maximize the diagnostic utility of existing data streams and enable early detection. One example is the diagnosis of left ventricular systolic dysfunction (LVSD) via ECGs. Traditionally, LVSD is diagnosed based on mechanical measurements of ejection fraction from the echocardiogram. Recently, AI has demonstrated an ability to interpret subtle electrical signals on an ECG that correlate to LVSD.⁵⁰ In addition, from a normal sinus rhythm ECG, it is possible for tools to discern latent signals of rhythm disturbances, such as AF or concealed long QT syndrome.^51,52 Similarly, AI can assist with the diagnosis of cardiac amyloidosis by inferring signals from echocardiogram videos, including point-of-care ultrasound, differentiating from other causes of increased left ventricular thickness as well as capturing preclinical progression, paving the way for widespread screening.53, 54, 55, 56

AI algorithms can also be used for “opportunistic sensing”; extracting additional, clinically useful information from a diagnostic modality that was not the primary reason the data were collected, without additional testing. For example, signatures of atherosclerotic cardiovascular disease risk are detectable on chest X-rays,⁵⁷ noncardiac computed tomography images,⁵⁸ mammography scans,^59,60 and retinal images.⁶¹ Some of the defining features, such as vascular calcification, are interpretable by expert human readers as well. However, the ability of AI tools to reliably identify on all testing, broadens the identification of diseases when competing diagnostic workflows may confound their detection in routine care.

Another “superhuman” capability of AI is the ability to forecast future clinical outcomes from a single, static diagnostic test. Convolutional neural network models trained on concurrent disease labels have displayed the ability to predict incident disease. For example, the application of an AI-ECG model trained to learn signatures of concurrent LVSD identified screen-positive individuals without HF at baseline linked with a >4-fold risk of developing LVSD in subsequent years.⁶² The same model also defined the risk of HF using a single baseline ECG, arguably identifying subtle structural anomalies that progress to HF in the future.⁵⁰ Similarly, an AI-Echo model trained to detect concurrent aortic stenosis could also predict the development and progression of severe aortic stenosis from a normal echocardiogram video.⁶³ In the future, AI-enhanced diagnostic modeling may harness multiple data sources and modalities in existing EHR data to uncover clinically actionable information.

Emerging AI model applications can also transform the whole temporal arc of a patient’s care into a digital narrative that can be used to infer a range of insights about their care and outcomes. An emerging class of AI models called transformers is particularly adept at encoding both structured EHR timelines and text into digital sequences that form foundational data layers for predictive risk models in an LHS.⁶⁴ Such an application enables robust longitudinal health modeling and can forecast disease states.⁶⁵ This formulation enables high-fidelity prediction of individualized disease trajectories, as well as supporting virtual trials and exploration of counterfactual interventions.⁶⁶ The framework also extends to specialized tasks, such as aligning structured EHR elements with other data sources, such as imaging, to capture a multimodal representation of disease states. For example, such a unified framework leveraged the EHR and cardiac magnetic resonance (CMR) to predict the risk of arrhythmic death in hypertrophic cardiomyopathy, with superior discrimination over established clinical guidelines.⁶⁷

Overall, AI models can support the realization of an LHS by converting existing clinical data points and free-text narratives into structured, forward-looking risk vectors.^10,27 In addition, AI-enhanced early detection and trajectory prediction have the potential to reconfigure care pathways from reactive to proactive management.^1,28,29 The combined impact may lead to improvements in important clinical endpoints—such as disability and mortality—as well as gains in workflow efficiency and patient experience.^2,16

Digitally augmented perception and action in a learning health system

An LHS can be conceptualized as having a “sensory” arm and a “motor” arm, which together translate information to improved care (Figure 2).⁶⁸ We discuss the sensing and nudging components of an LHS in detail below.

Figure 2.

Digitally Augmented Perception, Action, and Cognition in a Learning Health System

AI-enabled capabilities have the potential to augment the sensory, motor, and cognitive aspects of an LHS. Perception expands sensing through: 1) opportunistic sensing from routine tests and reports; 2) remote monitoring via wearables and implanted sensors that stream physiologic data; and 3) vocal biomarkers and ambient listening that convert clinician–patient dialogue into structured information. Action translates signals into decisions via: 1) digital nudges that support guideline-concordant therapy and quality improvement; 2) risk alerts that surface early deterioration to care teams; and 3) diagnostic dialogues that return individualized, evidence-grounded recommendations. Cognition closes the loop by: 1) simulation with digital twins for procedural planning and forecasting; 2) LLM-assisted trial recruitment; and 3) trial emulation and adaptive enrichment using real-world data to generate timely evidence.

Perception

Sensing in an LHS represents continuously capturing expansive, multimodal data from the health care environment.^11,49,69 AI-enhanced cardiovascular innovations can broaden data streams for richer perception.⁶⁸

An emerging sensory layer of an LHS combines voice recording capabilities with generative AI to streamline the workload of medical documentation.¹⁸ Ambient listening tools deploy LLMs on real-time clinician–patient dialogue to generate context-aware clinical notes and surface orders discussed for rapid verification.⁷⁰ Their use shortens note-completion time, lessening time spent after-hours on documentation, and enables physicians to focus on the uniquely human elements of their profession.⁴⁸ Consequently, ambient AI scribes have been shown to improve clinician efficiency and reduce administrative burden.^39,71 In addition, AI-enhanced speech analysis can also be used as a “vocal biomarker”. For example, classifying patients with HF as “wet” or “dry” by identifying subtle voice alterations linked to pulmonary fluid overload⁷² and assessing the risk of HF hospitalization and mortality.⁷³

AI-enabled extraction of data from wearable devices enables home-based assessment, bypassing the traditional framework of data collection at clinical encounters.^10,27,29,74 The use of the single-lead ECG capability of wearable devices, such as the Apple Watch or Fitbit, paired with AI facilitates the detection of cardiovascular disease.75, 76, 77 Moreover, multimodal AI models can distil rich patterns in time series data to infer advanced physiological metrics that presently require specialized laboratory testing. AI-enhanced wearable sensors can be used to infer peak oxygen uptake⁷⁸ or pulmonary capillary wedge pressure,^79,80 enabling assessment in the community. A nationally representative survey of U.S. adults found that most would be willing to share wearable data with clinicians.⁸¹ On the other hand, recent qualitative work has shown that such data streams can be converted into commercial assets with little patient involvement.⁸² Scaling up to enable population health screening for cardiovascular disease in the community will require thoughtful governance.⁸³

Remote patient monitoring using implanted devices is another way that the uptake of digital health technologies grows the sensory capabilities of an LHS.^10,27,29,84 In recipients of implantable cardioverter defibrillators, AI-enabled analysis of continuous data revealed a sedentary, poor-sleep profile linked with significantly higher rates of malignant ventricular arrhythmias.⁸⁵ Furthermore, a novel implantable sensor—measuring pulmonary artery pressure—significantly reduced HF hospitalizations by streaming daily hemodynamic data to clinicians.^86,87 The use of remote patient monitoring has increased ten-fold among Medicare enrollees between 2019 and 2022,⁸⁸ opening new avenues for timely clinician intervention while deepening patients’ engagement in their care.^10,27

Action

The use of AI tools to enhance precise interventions can be conceptualized as a “motor” arm of an LHS,⁶⁸ translating validated evidence into action via decision support or quality improvement.^11,49,69 For example, AI can detect early signals of impending clinical deterioration, then trigger an alert as a clinical decision support tool.¹⁰ In a pragmatic, single-blind randomized clinical trial (RCT) conducted in emergency and inpatient departments, mortality risk inferred using an AI application for ECGs was associated with reductions in 90-day all-cause mortality when the AI risk prediction was visible to the treating clinician. Exploratory analyses attributed this benefit to earlier escalation of care: higher rates of admission to intensive care units, amiodarone use, and evaluation with echocardiography or N-terminal pro-B-type natriuretic peptide testing.⁸⁹ In a pragmatic cluster-RCT spanning acute and intensive care units across 2 health systems, hourly AI alerts—derived from nursing-surveillance metadata—reduced in-hospital mortality when surfaced without interrupting workflows in the EHR to the care team. Further exploratory analyses suggest the survival benefit arose from earlier interventions for AI-detected high-risk patients.⁹⁰

A “nudge” is an implementation strategy that influences behavior predictably without restricting choice.⁹¹ In cardiovascular care, electronic “nudges” use behavioral economic principles to steer physician or patient choices toward optimal care, for example, by defaulting to guideline-directed medical therapy (GDMT) in order sets.⁹² However, across multiple randomized trials and registries, embedding nudges into clinician decision support or patient messaging has achieved only modest gains in achieving uptake of optimal therapies and improving medication adherence in cardiovascular care.93, 94, 95, 96, 97, 98 There needs to be a focus on improving the quality and appropriateness of nudges to improve their impact.⁹⁹ Individualized nudges could be supported by an AI-driven framework identifying the reasons for GDMT underutilization. For example, in patients with AF, a RAG-LLM can calculate stroke and bleeding risk scores from structured and unstructured EHR more rapidly and reliably than traditional approaches, as well as identifying reasons for nonadherence, including patient preferences or contraindications.⁴⁰ Similarly, an AI-enabled opportunistic sensing approach can be used to improve GDMT rates in hypertension. By scanning echocardiogram reports, an NLP model can identify mentions of left-ventricular hypertrophy in patients not yet prescribed an antihypertensive agent—even when hypertension management was not the focus of the study or care setting—to then automatically notify clinicians and prompt therapy initiation.¹⁰⁰

AI-enhanced analysis can also interpret continuous, multimodal physiologic data streams in real-time, enabling feedback loops.¹⁰ In a 1-year randomized trial of adults with both hypertension and type 2 diabetes, an AI algorithm that analyzed daily home blood pressure uploads together with continuous glucose monitor data led to hypertension remission (<140/90 mm Hg and no antihypertensive medication) in 50% of participants vs 0% under usual care.¹⁰¹ In the future, a clinician may be able to engage in an AI-enabled diagnostic dialogue, querying an LLM with comprehensive access to multimodal investigation reports, that return individualized recommendations and subspecialist-level differential diagnoses and management plans.¹⁰²

Overall, AI can generate precise and individualized electronic nudges, translating new evidence into action in real-time. Widespread implementation will require clear guidance around clinical actionability¹⁶ and consideration of the lead-time over which the AI model is evaluated.¹⁰³

Imaginative AI to forecast care and close the learning loop

Clinicians routinely imagine clinical trajectories, visualize complex procedures and reason through the evidence bases for competing therapies. Emerging AI approaches are poised to amplify these higher-order cognitive tasks, delivering high-fidelity digital simulations and rapid evidence generation across the LHS (Figure 2).

Simulation

The continuous improvement of care within an LHS can be advanced through digital twins, which are virtual representations of physical subjects that use real-time data to mirror the world experienced by the subject. In health care, digital twins can exist at various levels, ranging from individual patients to entire hospital systems, to simulate clinical scenarios, to inform medical decision-making, and forecast outcomes in real-time for the subject.^104,105 For example, an abnormality is detected on a patient’s wearable device and fed into the patient’s digital twin that is stored in an LHS. The digital twin then simulates possible outcomes from interventions and alerts the patient’s clinician of the event and their recommendation.¹⁰⁶ Notably, the substantial computational power required to develop digital twins—and the associated cost—must be balanced against equity as a key component of an LHS.

Applications of cardiovascular digital twins include simulation for procedural planning or guidance during the procedures. For instance, in AF ablation procedures, a digital twin model leverages CMR images to generate a 3D reconstruction of the atria, predefining fibrotic areas as targets for ablation, avoiding lengthy electrical mapping.¹⁰⁷ In a randomized trial, combining AI-guided target ablation with standard pulmonary vein isolation achieved higher 1-year freedom from documented AF than pulmonary vein isolation alone.¹⁰⁸ Digital twins can also be used to provide intraprocedural guidance for coronary angiography.¹⁰⁹ Real-time guidance of percutaneous coronary intervention is delivered by superimposing a coronary roadmap onto live fluoroscopy images, reducing contrast volume, radiation dose, and procedure time.¹¹⁰ Another intraprocedural digital twin model uses angiographic images to simulate the physiological results of stent placement faster than measurement of fractional flow reserve.¹¹¹

The ability of digital twins to simulate multiple patient characteristics in parallel permits granular and personalized forecasting of disease progression and modification.¹¹² In patients with a myocardial infarction, construction of 3D heart models from CMR images enables estimation of arrhythmia risk, potentially reducing the rate of sudden cardiac death and also avoiding unnecessary implantable cardioverter defibrillator insertions.¹¹³ In diabetes care, a digital twin modeling glucose dynamics to automatically titrate insulin delivery in patients with type 1 diabetes reached its primary outcome of improved time in glucose range.¹¹⁴

Digital twins typically use multimodal data available in an LHS, with substantial complexity involved in model building and deriving individualized parameters. In cardiovascular care, for example, digital twins mechanistically model processes from cells to organ systems. Accordingly, most models to date have focused on single organ systems and have been limited to a few dozen individuals. Recently, neural networks and Gaussian process emulation have reduced the computational space to derive individual parameters, allowing for the creation of digital twins of a large cohort of participants.¹¹⁵ Although a majority of applications have focused on individual patients, system-level models are poised to have a major impact on health care workflows, efficiency, and patient outcomes.¹⁰⁶ For example, a nation-wide digital twin of emergency call centers simulated call center distribution, cellular service availability, and call volume to optimize the service quality of response calls.¹¹⁶

Evidence generation

The application of generative AI to an LHS can also contribute to evidence generation through enhancing recruitment to clinical trials and emulating treatment effects in real-world populations. Historically, identifying eligible patients for clinical trials has been inefficient and costly, often relying on patient referrals from clinicians who may not be consistently aware or motivated.¹⁰ For example, a national registry found that <5% of eligible patients were enrolled in acute coronary syndrome trials.¹¹⁷

An LHS can improve inefficiencies in trial recruitment by leveraging EHRs to streamline patient identification and trial execution. LLM-based tools can facilitate the rapid screening of the EHR to identify eligible participants than manual review.118, 119, 120 Trialists could also foster uptake by applying less restrictive eligibility criteria, prospectively embedding implementation science principles of stakeholder engagement and workflow integration within the trial design.¹²¹ Indeed, AI-driven emulation of oncology trials using real-world EHR data has demonstrated that many eligibility rules can be safely relaxed, doubling the number of eligible patients while preserving estimated treatment effects.¹²² AI tools can also be used for more intelligent trial recruitment. For example, identifying potential participants with a higher mortality risk helps identify a population more likely to reach a mortality endpoint.¹²³ Alternatively, a phenomapping approach embedded in an adaptive trial design learns responder signatures at interim analyses to subsequently enroll participants conditioned on predicted benefit, reducing the number of participants required to achieve the same treatment effect.¹²⁴ Lastly, AI can also be used to augment clinical trials with in silico control populations.¹²⁵ An LHS can provide the patient population for these innovations and, in turn, learn from their findings.

Given the cost and duration of RCTs, trial emulation in observational EHR data sets represents the next frontier for evidence generation.¹²⁶ Targeted trial emulation takes advantage of the large sample sizes and rich longitudinal data available within an LHS, using advanced statistical techniques to balance confounders between treatment arms and estimate treatment effects.127, 128, 129

Finally, emerging applications of AI can enable inference from RCTs targeted for specific populations. Notably, RCT-measured treatment effects often generalize imperfectly to heterogeneous real-world populations, likely due to the under-representation of demographic groups enrolled in RCTs.^130,131 Digital twins provide an AI-enabled way to perform trials in hard-to-enroll populations—such as children—by supplying virtual control arms and reducing exposure, cost, and time while preserving inferential power.¹³² Another reason for the limited generalizability of RCT treatment effects may be greater prognostic heterogeneity among real-world patients.¹³³ Generative adversarial models such as RCT-Twin-GAN learn EHR covariate distributions to create a digital twin of the original trial, translating its treatment effects to real-world patients while preserving the estimated effect size.¹³⁴ These synthetic cohorts are abstractions—performing model-based augmentation to reweight and translate trial effects under explicit assumptions to a specified target population—but provide some representation where none exists. Taken together, AI approaches may close the learning loop using newly generated data to refine and improve future care.^2,49 Although the use of AI for evidence generation remains in its early stages, with limited impact on guidelines and clinical care so far, continued methodological research and application to important clinical questions is critical to realizing its full potential.

Guardrails against unwanted learning: maintaining fairness, privacy, and sustainability

AI algorithms can learn unintended patterns potentially leading to harm. For example, embedding systemic biases present in training data, enabling patient reidentification that compromises privacy, and drifting toward artifactual signals after deployment. In this section, we discuss these risks and outline guardrails for fair, privacy-preserving, and sustainable AI implementation (Figure 3).

Figure 3.

A Roadmap for Implementing “AI in Action”

A roadmap for “AI in action”, transforming cardiovascular health systems from sparse, reactive, hospital-centric episodes to a multimodal, connected learning ecosystem that continuously measures, learns, and improves cardiovascular outcomes. Guardrails ensure models learn the right signals and remain safe over time. Operationalization requires AI models to be generalizable, interoperable, and usable. Diffusion will require explainability, thoughtful regulation, and engaged stakeholders for scaling of responsible use. Together, these protections support the implementation of trustworthy AI within a learning health system, toward system-wide adoption and sustained improvement.

Fairness

A core principle for trustworthy AI is fairness, which mandates that AI tools maintain consistent performance across population subgroups.¹⁶ However, EHR data are inconsistent, incomplete, and prone to biases from underlying systemic issues.¹³⁵ These stem from nonrepresentative sampling and systematically poorer data quality for historically oppressed and excluded groups,^30,136 driven by inequitable health care access, socioeconomic disparities, and related structural factors.^137,138 Moreover, most AI models are developed to optimize aggregate performance, so they can appear accurate globally, but still perform poorly for under-represented subgroups, further perpetuating health disparities.¹³⁹ Therefore, it is essential to incorporate health equity frameworks to address this issue. Ideally, this should be addressed upstream by training AI tools on representative real-world data that adequately capture variations encountered in clinical practice.¹⁶ Practical steps will include lowering access barriers to health systems where most data are collected, broadening data collection to alternative sources such as wearables, and addressing mistrust about how health data are used for care and research.

Algorithmic bias is a major obstacle to the implementation of fair medical AI. One issue is “demographic shortcutting,” where deep learning models can base predictions on confounders, such as age, race, or ethnicity, instead of using task-relevant clinical features.¹⁴⁰ The algorithmic encoding of demographic attributes produces quantifiable unfairness, for example, a higher false negative rate in certain age groups, reflecting substantial underdiagnosis.¹⁴¹ The reliance on demographic shortcuts may be suppressed during model training by specific technical adaptations. These include adding an extra output layer to predict demographic attributes and applying either an adversarial penalty¹⁴¹ or gradient reversal¹⁴² to negate demographic prediction while optimizing the model backbone. Post-training calibration of model thresholds can further equalize error rates across demographic groups.¹⁴¹ Optimal approaches account for bias remain uncertain and likely will vary across data sources and use cases.

The adoption of generative AI in an LHS may present an additional form of algorithmic bias. LLMs were presented with a series of emergency department cases, each presented in 32 variations with socioeconomic demographic changes, but clinical information held constant. Cases labeled as Black, unhoused, or LGBTQIA+ were more often routed to urgent care or mental health assessment than controls. In addition, there was an inflation of advanced imaging referrals for high-income labels, whereas low-income cases were steered toward basic or no testing. These disparities were not supported by guidelines, indicating algorithmic bias driven by demographic proxies that reflect structural inequities rather than clinical need.¹⁴³ Future LLM workflows may address this via grounding—tying each AI model output back to verifiable evidence, such as radiology images, text notes, or clinical guidelines, allowing users to trace the provenance of the information¹⁸—and guardrails—layered controls designed to detect or block unsafe or ungrounded outputs from AI systems—forming a critical safety net.⁴⁷

Biases with diagnostic labels are another key consideration.¹⁴⁴ A widely used care-management algorithm meant to flag patients with complex needs was found to exhibit significant racial bias.¹⁴⁵ Instead of identifying clinical need, it had learned to optimize for future health care costs as a proxy label, therefore systematically underestimating risk for Black patients, who often incur lower costs due to access and insurance barriers. Emerging solutions include using direct health measures—laboratory values, diagnoses, and clinical endpoints—as AI model targets rather than proxy labels that can embed structural equities.¹⁴⁶ Fairness can also be explicitly assessed using frameworks such as probability-based checks; clinically meaningful, decision-analytic metrics is another strategy.¹⁴⁷ Implementation science methods then provide guardrails for deployment, ensuring that the benefits of AI are equitably distributed across all patient populations and health care settings, especially in under-resourced areas.¹³

Data privacy

Public trust is essential for the successful implementation and sustained operation of an LHS, particularly in the AI era.^2,148 This depends on strong data privacy and security frameworks. However, AI algorithms are data hungry. Richard Sutton, a founding father of reinforcement learning, wrote in 2019 of the “bitter lesson” made evident from 7 decades of AI research: expertly crafted, domain-specific models are outperformed over the long run by general-purpose algorithms, which simply harness more compute and data.¹⁴⁹ AI scaling laws empirically demonstrate Sutton’s thesis, showing that model performance approximates power-law gains as the volume of training data and compute grow.¹⁵⁰ Data scaling is the clear trajectory for medical AI,¹⁵¹ as reflected by a recent report of a generative model granted access to EHR data from 57 million patients.¹⁵²

Data privacy approaches based on simple deidentification are fragile, as linking “anonymized” health records with basic demographic information is sufficient to reidentify individuals.^153,154 Federated learning circumvents this by sharing model parameters across sites, rather than raw data.¹⁵⁵ Federated models may be augmented with differential privacy, a technique that injects carefully calibrated noise so that the presence or absence of any single patient’s data cannot be inferred.¹⁵⁶ Another solution may be provided by generating synthetic data using generative adversarial networks or probabilistic graphical models to preserve real distributions and dependencies.^157,158 Taken together, these AI-driven solutions to data privacy enable model fitting while minimizing patient identifiability and privacy exposure.

Adaptive learning

Unlike traditional models—which are static, rule-based, and require manual updates—AI models can adapt to maintain or improve performance postdeployment.^2,26,29,49 As patient demographics, clinical workflows, and data-capture practices drift over time, the population distribution that was used for model development does not match reality, which can result in the degradation of model performance and risk harm.¹⁸ Although the adaptive property of AI modeling is well-suited to an LHS principle of continuous learning,^148,159 there are a number of challenges. Sensing the need for a model update requires differentiating real and artifactual performance degradation, with the latter arising in the setting of the model appropriately modifying care pathways. This may require the development of novel counterfactual analysis methods to assess performance accurately.¹⁶⁰ Another challenge is deciding when to implement model updates. Implementation science outlines the need for continuous monitoring, periodic auditing, and iterative updating of AI systems postdeployment.¹⁷ For example, the “tight-loose-tight” framework, adapted for health care AI, explicitly includes a “monitor and report outcomes” phase to ascertain if the intended results are achieved.¹⁸ An alternative to this “freeze-the-model” approach may be a “dynamic deployment” framework, using continuous assessment of local performance.¹⁶¹ Lastly, further research is required to determine how to implement updates. A simulation study indicated that, once an intervention is triggered, retraining or codeploying additional models degraded model specificity at fixed sensitivity, meaning no universal update rule existed.¹⁶² The complete lifecycle of an AI model must likely be tracked to sustain performance, as well as ensuring traceability and trust.¹⁶

Operationalizing AI in the learning health system

A spectrum of implementation challenges shapes whether AI will deliver value in cardiovascular care. To operationalize AI, the central questions are as follows: will the model generalize? can it operate with local data and systems? and will clinicians adopt it? This section outlines AI universality, interoperability, and usability as determinants of success in an LHS (Figure 3).

Universality

The generalizability of AI models to new populations, clinical sites, and equipment remains a key challenge for integration into an LHS.² Implementation science calls for robust, prospective validation in diverse, real-world clinical workflows. However, most AI algorithms fall short of this standard, despite showing promise in retrospective studies.¹⁶³ Recent progress includes external benchmarking of AI-ECG models for LVSD detection¹⁶⁴ and AI-Echo models for amyloidosis.¹⁶⁵ However, more evidence to support the effectiveness, safety, and value of AI applications is needed, while ensuring that evidentiary standards for these applications are comparable to other screening or intervention tools. For example, in an analytical model study, initiating anticoagulation in patients with wearable-detected subclinical AF led to negligible benefit and substantially higher bleeding risk.¹⁶⁶ Evaluations of costs and potential harms of this kind are essential for stakeholder trust. AI tools that fail to meet prespecified performance and utility thresholds should be retired.

AI systems can be exquisitely sensitive to modest, often imperceptible variations in input data. Accordingly, the intentional adaptation of models to account for local context may be a key feature of effective AI performance, rather than the pursuit of a single static “generalized” model. Input variations can arise from a range of factors, such as differences in measurement devices or data acquisition protocols^167,168 or targeted adversarial attacks.¹⁶⁹ One solution is to evaluate local clinical validity, then recalibrate or fine-tune the model if needed.¹⁶ This approach has been applied to locally calibrate AI-ECG models for hypertrophic cardiomyopathy¹⁷⁰ and LVSD.¹⁷¹ Another solution is equipment-specific fine-tuning. For example, an AI-Echo model for detecting aortic stenosis preserved accuracy on CMR by applying a cross-modal adapter layer that reshaped CMR videos into the model’s native echo format.⁶³ Lastly, preclinical silent trials, which run in the background without influencing clinical care, provide a novel way to evaluate AI tools in new populations. These trials have been endorsed by ethical AI implementation frameworks² and may assist with successful downstream implementation.¹⁷²

Interoperability

Interoperability remains a major challenge to the broad implementation of AI in cardiovascular care.¹¹ Disparate systems store data in incompatible formats, structures, and vocabularies, so a digital tool developed in one system struggles to understand the information received from another.¹⁷³ The Fast Healthcare Interoperability Resources standard stresses the need for rigorous syntactic and semantic definitions to promote the exchange of health care-related data.¹⁷⁴ These include using standard languages for medical ontologies such as SNOMED (Systematized Nomenclature of Medicine) codes, common data models like OMOP (Observational Medical Outcomes Partnership), and interoperability protocols including DICOM (Digital Imaging and Communications in Medicine) and HL7 FHIR (Health Level Seven Fast Healthcare Interoperability Resources).¹⁶ In parallel, emerging “zero-shot” approaches, in which a transformer model can adapt to a new hospital’s data schema with minimal additional tuning, may provide an AI-based solution to data interoperability.¹⁷⁵ Resolving fragmentation between noninteroperable systems could yield higher-quality more representative training data²⁷ and accelerate the knowledge generation cycle¹⁰ across an LHS.

Usability

Effective AI deployment critically depends on usability and workflow integration.^2,11 The optimization of these elements helps ensure that new technologies and evidence-based practices are adopted and demonstrably improve both patient care and clinician experience.^16,29 Implementation science and human-computer interaction methods, such as rapid-cycle A/B testing, provide structured ways to evaluate usability and iteratively refine interventions.¹⁷⁶ Implementation studies show that real-world pressures—such as time pressure, alert fatigue, and data overload—can erode the usability of a system.^177,178 User feedback is therefore vital for refining workflow integration, ensuring that critical information is surfaced and delivered to patients and clinicians in their preferred format.^179,180 Multifaceted implementation strategies that combine technology updates, staff training, and policy changes can also accelerate the adoption of evidence-based practices.¹⁸¹

AI introduces specific challenges, notably the need to reconcile algorithmic recommendations with clinician judgment. For example, in an implementation across 3 emergency departments, an AI-informed triage assistant increased sensitivity for identifying patients requiring critical care and improved patient flow. However, nurse agreement with the AI recommendation varied—high-agreement nurses outperformed the AI alone, whereas low-agreement nurses underperformed it—underscoring the need to calibrate the balance between artificial and human intelligence.¹⁸² Interviews with physicians and nurses using an AI-alert system for sepsis suggest a partnership model may be a good way to build trust over time while supporting clinician autonomy.¹⁸³

Accelerating the diffusion of AI into clinical practice

For the diffusion of AI into cardiovascular care pathways, health systems must combine early ongoing stakeholder engagement; transparency in AI-guided decisions; and governance aligned with evolving frameworks for learning algorithms. Together, these pillars foster a culture of shared responsibility that enables trustworthy, adaptive implementation (Figure 3).

Explainability

Many AI models function as “black boxes”, offering little insight into how they reach a conclusion. This lack of transparency stymies adoption in evidence-based medicine, where decisions can carry life-or-death consequences. Consequently, explainability—the ability of a system to reveal clinically meaningful rationale for its predictions—has become a key research focus.¹⁶

Explainability approaches—such as SHapley Additive exPlanations, Local Interpretable Model-agnostic Explanations, and Gradient-weighted Class Activation Mapping—highlight the key input elements driving model predictions. For example, in AI-enabled analyses of clinical notes, these explainability methods can highlight the text features which are the most influential in the model’s decision-making.⁴⁰ In AI-ECG applications, they can offer insight into which pixels or waveform segments are used by algorithms to detect disease.^50,184 Although current explainability methods can aid system-level audit, they may be unreliable and can mislead at the patient level.¹⁸⁵

As these approaches mature, they have the potential to create transparent feedback loops that improves clinician confidence and potentially even sharpens human diagnostic skills. Exposure to AI-ECG mortality scores improved cardiologists’ ability to identify high-risk ECGs.¹⁸⁶ In another study, researchers clustered SHapley Additive exPlanations values for an LVSD model to derive 6 human-readable ECG patterns. After a short explainer video, this approach also raised cardiologists’ reading accuracy.¹⁸⁷ Lastly, generative AI provided an interactive tool, synthesizing counterfactual ECGs to simulate the textbook waveform and rhythm changes associated with hyperkalemia and AF, thereby enhancing clinician trust.¹⁸⁸

Regulatory and policy

The regulatory landscape for AI in health care is evolving rapidly.¹⁸⁹ Traditional policy frameworks—built for static medical devices—are being reimagined to accommodate the continuous learning cycles enabled by AI tools embedded in an LHS.¹⁹⁰ The FDA proposed Software as a Medical Device guidance¹⁹¹ defines an iterative evidence cycle: 1) valid clinical association between the software output and target condition, 2) technical validation proving the software processes inputs accurately and reliably, and 3) clinical validation confirming the output improves care in the intended population. Pertaining to AI-enabled software, the Software as a Medical Device framework emphasizes oversight of the entire lifecycle. This includes the need for postdeployment performance monitoring for drift and the use of a predetermined change control plan to streamline future model updates without necessitating full re-review. The National Academy of Medicine’s AI Code of Conduct reinforces this emphasis on continuous evaluation.¹⁸ However, adaptive accountability appears to remain the key bottleneck framework for safely implementing AI tools. A recent review of nearly 1000 FDA-cleared AI medical devices found that most postmarket recalls stemmed from software-performance defects and labeling errors, underscoring the need for continuous real-world surveillance to trigger timely model updates.⁷ Building empirical evidence may prove essential to shape AI governance frameworks, including policies that best harness commercial resources and expertise while maintaining safeguards to protect data privacy and clinician autonomy.

Stakeholder engagement

Successfully integrating AI into an LHS depends on meaningful engagement with stakeholders involved in care delivery and improvement processes.192, 193, 194 Key stakeholders—including clinicians, patients, administrators, payers, information technology vendors, regulators, and researchers—should collaboratively shape the responsible implementation of AI across health systems.¹⁹⁵

Clinicians express concerns around over-reliance on AI systems, including ownership over decision-making, “automation bias”, and deskilling.^183,196,197 A “human-in-the-loop” system, relying on expert human input to adjudicate the AI model, could help mitigate these concerns by surfacing high-stakes decisions for clinician review and quantifying model uncertainty.¹⁹⁸ Scaling human-in-the-loop workflows, while keeping clinician workload manageable, may require automation of low-risk cases and defining clear escalation pathways. Patients are wary of the depersonalization of care interactions and data privacy risks.199, 200, 201 Conversely, patients are more accepting of transparent, clinician-supervised health care AI.²⁰² Interestingly, AI tools can also be used to facilitate patient engagement, for example, by using LLMs to identify care priorities.²⁰³ Health care leaders and researchers focus on the organizational changes required for widespread AI adoption, including infrastructure readiness, workforce training, and ethical governance.^204,205 Cost is another concern requiring consideration of funding, reimbursement models, and resource allocation.^2,13,159 The concern about AI replacing the clinical workforce should be weighed against AI’s potential to democratize care amid physician shortages in the United States and worldwide.²⁰⁶

Stakeholder engagement should be integrated at every stage of the AI tool lifecycle.¹⁴ At the design stage, stakeholders can thoughtfully codesign AI applications in alignment with local priorities.^196,207 At the implementation stage, the design of AI clinical workflows can be iteratively refined according to user feedback and real-world constraints.¹⁸³ Across deployment and evaluation phases, ongoing collaboration between AI developers, clinicians, and patients is essential to ensure models remain performant, safe, and equitable as workflows and data evolve.²⁰⁸

AI in action

This state-of-the-art review outlines how AI can serve as a copilot for cardiovascular care within an LHS—from the automation of care outcome measures and diagnostic reporting tasks to advances in disease phenotyping and risk forecasting. The operationalization and diffusion of AI tools within an LHS requires the uniform application of core principles of evidence-based practice—explicitly addressing a range of logistical, system, and human factors—and guardrails to ensure efficient, safe, and effective care. Successful translation of AI into action demands the alignment of augmented capabilities with continuous learning via feedback loops to ensure sustained improvements in patient outcomes, health care quality, and equity.

Funding support and author disclosures

Dr Biswas is a co-inventor on U.S. patent applications 2025/0084489 and 2022/0136063. Dr Ogunniyi has received institutional research support grants from AstraZeneca, Boehringer Ingelheim, Cardurion Pharmaceuticals, and Pfizer, outside the submitted work. Dr Thomas Maddox reports serving as a consultant to Elion (advisory board member). Dr Ahmad has received research support from Pfizer, Atman Health, Tempus, AstraZeneca, and Abiomed; consulting fees from Alnylam Pharmaceuticals; and speaking honoraria from AstraZeneca and Alnylam Pharmaceuticals. Dr Khera is supported by the National Institutes of Health (R01AG089981, R01HL167858, K23HL153775) and the Doris Duke Charitable Foundation (Award 2022060), with additional support from the Blavatnik Foundation through the Blavatnik Fund for Innovation at Yale; he is an Associate Editor of JAMA; through Yale, he has received research support from Bristol Myers Squibb, BridgeBio, and Novo Nordisk; he is a co-inventor on patent applications WO2023230345A1, US20220336048A1, 63/484,426, 63/508,315, 63/580,137, 63/606,203, 63/619,241, and 63/562,335, and 63/346,610; and he is a cofounder of Evidence2Health, and Ensight-AI. The funders had no role in the preparation, review, or approval of the paper, and decision to submit the paper for publication. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.