Transforming Cardiovascular Care with Artificial Intelligence: From Discovery to Practice

2a. Innovations in disease diagnosis

Traditionally, disease diagnosis has relied on human-led interpretation of multimodal inputs in history, examination findings, imaging, and test interpretation. Given their computational simplicity, the earliest studies in AI and machine learning (ML) focused on developing tools that relied only on the structured, tabular data derived from these care encounters. Their focus was the standardized and efficient screening of cardiovascular disease under the premise that several ML algorithms capture non-linear relationships and, therefore, may be more effective than traditional linear models in fitting real-world data where complex relationships between different exposures and outcomes exist. In one example, this was realized in an algorithm for diagnosing coronary artery disease from electronic health record (EHR) datasets.⁵ However, head-to-head studies that have compared the performance of non-linear models developed in such tabular datasets have not found them superior to linear models developed on the same inputs.⁶

Instead, AI might offer the greatest value by identifying sources of phenotypic variation and risk in complex signals rather than tabular data. With advances in deep learning models, such as those based on convolutional neural networks and transformers, AI algorithms can process raw, unstructured biometric signals and images, learning new representations from these data beyond what is manually inferred and encoded. Perhaps the most studied example is the AI-guided interpretation of 12-lead electrocardiograms (ECGs). Several studies demonstrate that AI-ECG can screen for the risk of paroxysmal atrial fibrillation using ECGs in sinus rhythm,⁷ but can also detect structural heart disease through electrocardiographic signatures that may precede overt myocardial function changes on echocardiography (Figure 2),⁸ including a range of cardiomyopathies such as left ventricular systolic dysfunction (LVSD),^9,10 hypertrophic cardiomyopathy (HCM),^11,12 transthyretin amyloid cardiomyopathy,^13,14 and even valvular heart disorders.¹⁵ These applications have been expanded to single-lead ECG acquisitions through AI-assisted stethoscopes, smartwatches, and portable devices that acquire fewer leads,^16–18 facilitating easy deployment, scalability, and accessibility at the point-of-care, for both chronic (i.e., LVSD)¹⁸ and acutely life-threatening conditions such as ST-elevation myocardial infarction.¹⁹

AI has also enabled better access to more advanced diagnostic tests that are limited by the need for expertise in acquisition and interpretation more than the equipment. For example, cardiac ultrasounds are increasingly obtainable on portable devices, but their acquisition until now has required specific training. Deep learning-enabled guidance systems provide explicit directions on the appropriate positioning to enable novices to acquire high-quality images.²⁰ Moreover, the interpretation of these videos has been automated by a series of tools to measure and classify observable diseases.^21,22 In addition, emerging tools augment the capacity of human readers, identifying individual disorders in complex disease categories,²³ defining the presence and degree of valvular disease without requiring advanced testing with Doppler,^24,25 and even interpreting prenatal screening ultrasounds for congenital heart disease.²⁶

AI further enables opportunistic screening of cardiovascular disease across modalities. Several studies now show that chest X-rays can be used to detect coronary atherosclerosis,²⁷ aortic stenosis,²⁸ and LVSD.²⁹ Estimation of LVSD may also be feasible from coronary angiography films, enabling assessments during emergent procedures.³⁰ The scalability of some of these algorithms may be augmented by their integration into standard screening protocols, such as by detecting coronary calcifications on low-dose computed tomography scans for lung cancer screening³¹ or breast arterial calcifications on mammography.³² Finally, automated pipelines may be integrated into routine cardiac diagnostics, such as coronary computed tomography angiography, to better quantify plaque burden³³ and provide insights into biological and inflammatory activity,³⁴ thus expanding the diagnostic yield of standard reporting.

Across all these applications, scrutiny is required to ensure the diagnostic signal learned is directly linked to the label of interest rather than a confounder.³⁵ Careful matching of cases and controls on a range of clinical and demographic factors may address measured confounders, but extensive testing across external and independent cohorts is essential to provide further reassurance about the generalizability of any given tool.

2b. Digital biomarkers of disease risk

Most AI developments for digital data streams in healthcare have focused on optimizing disease diagnosis or prognosticating outcomes. However, another key domain is the discovery of new markers for those at risk of developing disease, which previously required detailed clinical evaluation, invasive blood-based biomarkers, or risk scores. The limitation of the current approach is the need for accurate assessment of the individual risk factors and calculating the risk in a busy clinical setting. Digital biomarkers can be passive and computable using existing or easily obtainable clinical data and promise to enable risk-informed care through deployments in existing data ecosystems. The relevant data streams are already available in many individuals before the onset of clinical disease or can be easily acquired.

Commercially available wearable devices are an ideal tool for disease prediction, as they are for disease screening, given their broad potential reach among healthy adults.³⁶ The metrics available on these devices, such as step counts, heart rate variability, and sleep quality, are strong predictors of cardiovascular health and outcomes, portending as much as a 2-fold variation in developing cardiovascular disease and premature death.^37,38 Advances in deep learning are now enabling an even more comprehensive use of these data and triangulating information across sources.³⁹

However, the selectivity among those who opt to use these devices and among those who use them consistently likely exerts a potent confounding effect.^40,41 The on-device availability of spot measures such as 1-lead ECG can overcome some of these predictive challenges with time-dependent data from wearables, especially with the emerging evidence of a broad suite of novel AI-ECG risk biomarkers. These include serving as a biomarker for accelerated aging as a digital surrogate of the biological age,^42,43 with individuals with ECG-defined biological age older than their actual age by 8–10 years having a 2-fold higher hazard of all-cause death than their counterparts.

There are also more specific risk markers that can be derived from ECGs with models developed to identify the cross-sectional presence of LVSD able also to identify those at 3–4-fold higher risk of developing LVSD in the future and those at risk of developing clinical heart failure.^9,10 More recent assessments also suggest that these risks may have therapeutic implications. Examples include stratifying the risk of cancer therapeutics-related cardiac dysfunction by deploying AI to baseline and on-treatment ECG studies.^44,45 Similarly, perioperative risk-defined AI-ECG among those undergoing non-cardiac surgery exceeds the predictive performance of established risk scores.⁴⁶

The rich predictive information is not restricted to ECGs, with chest X-rays able to risk stratify patients for coronary artery disease and estimate the 10-year risk of major adverse cardiovascular events.^27,47 More recently, deep learning has identified predictive cues in clinical domains where clinically relevant features were not interpretable through standard algorithms. Specifically, a new tool called the Digital Aortic Stenosis Severity Index (DASSi) was able to detect both the cross-sectional presence of severe aortic stenosis on a single cardiac ultrasound view but also identify those without aortic stenosis at risk of developing the disease and those with mild disease at risk of rapid progression.^25,48 The same tool was also extensible to other cardiac video-based data formats, such as cine cardiac magnetic resonance imaging.

Overall, since our clinical practice is inherently multimodal, AI will increasingly integrate multimodal elements to refine cardiovascular risk stratification (Figure 3). This includes composites of tabular EHR data with imaging,⁴⁹ genetic determinants integrated with dynamic cardiometabolic phenotypes,⁵⁰ environmental parameters (“exposome”),⁵¹ including fine particulate matter data⁵² and images and videos extracted from the patient’s local environment,⁵³ as well as biomarkers derived from natural language processing (NLP) and large language models (LLM) that can ingest clinical notes.⁵⁴ With the rapid expansion of informatics pipelines that can process this information, universal risk calculators may eventually be replaced by locally adapted and continually validated tools that account for the unique characteristics of local populations and communities.⁵⁵ Overall, from the simplest to the most complex data streams, AI-based digital biomarkers are enabling a new frontier in cardiovascular risk stratification that is more scalable and broadly usable.⁵⁶

2c. Novel approaches to disease prognostication

AI-driven tools are also enhancing another domain – defining a person’s unique disease trajectory and risk of adverse outcomes. Developing precise risk stratification tools is essential to guide patient counseling and risk-modifying therapies but is inherently limited by the stochastic nature of clinical events and their uncertainty. Traditionally, these stratification schemes have relied on limited numbers of simplified yet interpretable, human-encoded variables,⁵⁷ such as age and comorbidities. However, these tend to ignore the complex heterogeneity and interactions seen in biological systems, potentially limiting their prognostic value. The advent of complex models that can account for the time-varying inputs in the EHR and documentation in clinical encounters comes closer to realizing the full potential of common data streams in the EHR.^58,59

AI enables the extension of prognostication to unstructured data, such as 12-lead ECGs,^43,60,61 chest X-rays,⁴⁷ and echocardiograms,⁶² which effectively predict cardiovascular events and all-cause mortality across myriad conditions. This is most likely driven by detecting structural or electrical abnormalities directly or indirectly related to established cardiovascular disease or markers associated with biological aging and systemic risk factors.³⁵

The prognosticating role of AI may also extend to enabling high-value care practices and, as discussed later, individualizing how therapies may affect outcomes.^63,64 For instance, despite the heterogeneous risk, primary prevention implantable cardioverter defibrillators (ICD) are recommended in a large subset of patients with heart failure. Through multimodal integration of imaging phenotypes and tabular characteristics, AI tools may provide actionable risk discrimination for arrhythmic versus non-arrhythmic mortality,⁶⁵ guiding personalized risk-benefit discussions. AI models are also improving upon traditional rule-based approaches in discriminating underlying rhythm, thus preventing inappropriate shocks.⁶⁶ Further insights into personalized disease trajectories may be provided by “digital twin” approaches that model the effects of varying population features and their simulated effects on outcomes.⁶⁷

Emerging AI tools will also enable optimal outcomes from therapeutic interventions. In procedural cardiovascular care, for example, atrial fibrillation ablation is aided by deep learning models that integrate intra-atrial and surface signals.⁶⁸ For matching patients to medications, AI phenotyping of baseline and follow-up 12-lead ECGs may assist in the safe initiation and titration of disease-modifying therapies, such as novel myosin inhibitors in hypertrophic cardiomyopathy.⁶⁹ Other models can allow safe antiarrhythmic initiation to predict QTc prolongation and safe use.⁷⁰

Therefore, AI’s role spans the gamut of describing the risk profile of the patient and their disease trajectory, providing actionable insights that may personalize both diagnostic and therapeutic management.

3a. Locally adapted, continually learning AI tools

While the current phase of AI model development has focused on building the best tools that retain average performance across deployments, key aspects of populations and disease may enable models to perform better if they are explicitly adapted to these unique care settings.^55,71 Deep learning models can be explicitly fine-tuned as they encounter more data and, therefore, can adapt to local populations. The key innovations that would enable this future are the ability to consistently define gold standard labels across individual sites and the continued ability to leverage them even after the model is deployed. This may also address population drifts over time or performance degradation in minority populations by exposing the models to populations in the real world that differ from the population where they are developed. This is particularly important in cases where the successful implementation of an AI system results in changes in its input features and target population that may impact its longitudinal performance.⁷²

Some key safeguards are that the data sources in new settings do not reflect a bias in care that is encoded into the models, and that the training process exposes AI algorithms to the full breadth of clinical phenotypes, especially among under-represented communities (e.g. different skin colors in dermatology),⁷³ while preventing them from making predictions through demographic confounders of disease.³⁵ However, significant resources will be needed to steward these AI models and to evaluate and adapt models to local populations.⁷⁴

3b. AI-augmented clinical encounters

AI at the point-of-care promises to accelerate and standardize tedious and time-consuming tasks, with the potential of maximizing direct patient care during clinical encounters. These now cover the full spectrum of a typical clinical encounter, from standardized examination and assessment to automated clinical documentation. Indeed, assistive AI technologies may automate the grading and characterization of cardiopulmonary auscultation,⁷⁶ and even augment point-of-care inference with the addition of single-lead ECG recordings. This paradigm has been illustrated in the screening of LVSD, both in a cohort of women at risk for peripartum cardiomyopathy in Nigeria,⁷⁷ and a select cohort of patients with suspected cardiomyopathy referred for echocardiography in London.¹⁷

Additionally, previously underutilized biometric signals may also provide diagnostic value, such as voice-based analytics to estimate changes in volume status,⁷⁸ sensor-based gait-tracking technologies to grade frailty,⁷⁹ as well as direct portable echocardiography assessment through simplified protocols,²⁰ which maintain robust performance for the screening of valvular and structural cardiomyopathies.^24,25 Concurrently, advances in NLP and LLMs have paved the path for AI solutions that can produce clinical summaries based on recordings of ambient conversations.⁸⁰

3c. AI to enable high-quality community care

The evolution of remote sensing technologies has enabled a new paradigm of continuous monitoring, thus offering real-world insights into the fluctuation and variation of patient-centered information between healthcare visits.⁸¹ This may overcome the limitations of traditional in-person visits that only provide objective metrics at a given time, precluding an assessment of the holistic health and wellness of patients during daily activities. Instead, through extended monitoring of patients in their familiar settings, these tools provide data that could inform risk assessment for adverse cardiovascular outcomes. The deployment of remote monitoring now includes voice-based technologies,⁷⁸ dielectric sensing tools,⁸² and image-based assessment of lower extremities or jugular venous pulsations,^83,84 in addition to information from current patient-facing devices such as those that leverage photoplethysmography, accelerometers, and on-device 1-lead ECG,^85–89 in providing a global, yet non-invasive assessment of the overall cardiovascular status.

These technologies may soon provide a non-invasive alternative to implantable monitoring⁹⁰ and a long-term solution to ambulatory blood pressure or rhythm monitoring.⁹¹ Furthermore, rapid advances in material engineering may boost the usability of these algorithms by embedding reliable sensors into everyday clothes (“smart clothing”),^92,93 and by generating wearables patches that can simplify the acquisition process for simple echocardiographic videos in the community (Figure 4).⁹⁴ The deployment and personalization of these mobile and wearable technologies can be further enhanced through just-in-time adaptive interventions that learn to adapt their notifications and triggers to an individual’s state to maximize their effectiveness,⁹⁵ and may be further boosted by representation and continual learning frameworks that will adapt them to specific use scenarios, such as in clinics or remote health settings.⁹⁶

3d. Democratizing global cardiovascular care with AI

The reliance of current diagnostic cascades on advanced diagnostic testing creates stark differences in the appropriate diagnosis and treatment in low-resource settings.⁹⁷ The timeliness of diagnosis, stage of disease, and the prognosis and outcomes of patients vary substantially across care settings, with the largest gaps observable in the underdeveloped and developing world.⁹⁸ Even tests that are more accessible, such as ECGs and chest X-rays, require access to trained experts, who are often overworked. In the future, the focus on diagnostic cascades will likely be on risk stratification with AI tools deployed for portable devices that enable untrained individuals to do scalable community-based screening.⁹⁹ Similarly, ECGs acquired as part of standard clinical care can have automated interpretation for both standard conduction and rhythm disorders,^18,100 but also a range of structural heart disorders,¹⁸ further enabling risk-informed care. Similarly, handheld ultrasound yielded by minimally trained individuals will allow a clinical-grade echocardiographic evaluation right in small health centers.²⁰ Therefore, AI may narrow the gap in access to high-quality diagnostic care, though only with dedicated effort.

4a. Addressing data privacy and security

AI algorithms, especially the recent generative AI models, require access to large volumes of sensitive, multimodal patient data, the collection, processing, and use of which poses significant security and privacy risks. De-identifying the data used to train AI models is challenging, as the ‘richness’ of the data means that even anonymized patterns can be re-identified by malicious actors if they gain access to training databases.¹⁰¹ Additionally, the methods of data exchange or data sharing used by the various actors in the AI development pipeline are vulnerable, increasing the risk of data breaches.¹⁰² Finally, models can be exploited once released. Generative AI models can, for example, be fed prompts that force the release of personal health information that can be used for re-identification or other malicious purposes if trained on clinical data with patient identifiers.

Governments and regulators are aware of the need to mitigate these risks, with privacy and security frequently featured in national AI policy documents. China’s Interim Administrative Measures for Generative AI services, for example, requires generative AI developers to register their models and submit them for security inspection;¹⁰³ the EU’s General Data Protection Regulation explicitly applies to AI systems; and the first principles of the UK’s proposed regulatory framework for AI is ‘safety, security, and robustness’.¹⁰⁴ In the US, many groups are getting involved, including the Health AI Partnership, Coalition for Health AI, the American Medical Association, the Joint Commission (Responsible Use of Health Data), the National Academy of Medicine, and many others.

Consequently, multiple technical solutions to the privacy and security risks posed by AI are available: model-to-data techniques, such as federated learning, where the AI model is sent to local datasets where it is trained ‘in-situ’ limiting the need to rely on risky data sharing methods; synthetic data noise can be added to datasets to mitigate the risk of re-identification or inference attacks;¹⁰⁵ and ‘white hat’ attacks such as model inversion attacks can be conducted to identify security weaknesses.¹⁰⁶ None of these techniques are entirely immune to attacks, and each has different strengths and weaknesses; which technique is the right or best solution depends on the specific use case of the AI in question and the sensitivity of the data involved.¹⁰⁷ Without clear solutions to mitigating data and privacy risks, these challenges risk hampering AI efforts.

4b. Working to achieve interoperability across health systems and communities at large

An essential element of any AI algorithm development is their scaling to clinical practice. However, most AI algorithms in published literature have no clear path to real-world care, and a vast majority are not even externally validated.^108,109 Though financial and regulatory considerations are often barriers, the major challenge is that the algorithms are often not designed considering their interoperability across institutions. For example, it can often be challenging to deploy a model that leverages EHR data outside the parent institution due to vast variations in types of EHRs. However, this is feasible, as demonstrated through the multi-site EHR-embedded tools in the COVID-19 pandemic.¹¹⁰ Interoperability standards under national policy seek to address this, in part, by instituting data standards such as FHIR and HL7 and a mandate to support access to these outputs from EHRs via accessible application programming interfaces.¹¹¹ The emergence of AI applications for EHR has put further pressure on these standards to be realized and instituted consistently.

Moreover, independent groups have championed common data models, such as the OMOP (Observational Medical Outcomes Partnership), that flexibly enable working across different data standards after mapping them to the same common formats, though admittedly do not yet fully capture unstructured data such as images.¹¹² Moreover, these data models also enable federated deployment of statistical or machine learning models without sharing data across institutions.^113,114 Similarly, other AI applications, including those meant for language, are more adaptable with the emergence of LLMs that can parse meaning in differently structured text.

Finally, AI models built for different cardiovascular modalities, such ECGs and imaging, are most scalable when they align with the prevailing data structure (such as ECG images vs signals),¹¹⁵ and aligned with standard data storage formats, such as Digital Imaging and Communications in Medicine (DICOM).¹¹⁶ The broad use of DICOM, a recognized data standard for medical imaging, has enabled AI for medical imaging to grow faster than other domains. Therefore, ensuring interoperable models requires the development of algorithms on interoperable data streams and would need to become a prerequisite for AI to be scalable and usable in clinical practice.

4c. Enabling appropriate regulatory controls in the era of foundation models

The development, deployment, and use of AI for health care is legally complex. First, AI designed to support diagnosis, prevention, monitoring, prognosis, or treatment should be considered a medical device and subject to regulatory control intended to ensure quality, safety, and efficacy.¹¹⁷ Yet, evolving medical device laws in the UK, EU, and the US all currently lack clarity - particularly with regard to evidence standards for AI – and do not cover generative AI. Consequently, the evidence base supporting AI interventions is weak. For example, in October 2023, of the 691 AI/ML-enabled medical devices approved by the FDA, just 9 (1.4%) had been tested in an interventional trial, and 96.7% had been approved via the 510(k) pathway,¹¹⁸ which typically does not require submission of clinical data. Second, there are questions regarding accountability and liability. It is not clear, for example, whether clinicians will be held liable in cases of missed or misdiagnosis involving a diagnostic algorithm used in a clinical setting, nor how the burden of proof would be met in such cases.¹¹⁹ Third, it is unknown how general, yet highly relevant, regulations relating to data protection and anti-discrimination will be adapted for AI.¹²⁰ In the EU and the UK, the legality of medical data processing depends on the context and intended purpose: using data for research requires explicit patient consent, whereas using data for direct care does not. AI blends these different use cases by, for example, providing diagnostic support for direct care and continuing to learn (‘research’),¹²¹ and its black-box nature makes obtaining specific and unambiguous consent difficult. Furthermore, EU and US anti-discrimination law is typically technology agnostic but only covers specific areas of activity, such as employment and education, rather than important facets, like bias among algorithms used in healthcare.¹²⁰ Finally, generative AI introduces a range of novel safety concerns related to, for example, its tendency to hallucinate false information that no existing laws are designed to cover.¹¹⁷ There are developments underway to overcome some of these regulatory limitations. The US Food and Drug Administration (FDA) has, for example, published an AI/ML-enabled medical device action plan,¹²² but the pace of regulatory change does not match the pace of the technology.

Moreover, given the resources required to train, validate, and test AI tools in specific settings, we anticipate significant ‘off-label’ use of AI in healthcare, that is, unapproved use of an approved model.¹²³ This paradigm is expected to become even more common in the near future with the arrival of task-agnostic foundation models. Since most existing regulatory frameworks require a defined intended use, the feasibility of regulating clinical foundation models likely requires the development of novel regulatory pathways to cover the intended and unintended uses of such models.

5a. AI for biological discoveries: mechanistic inference and drug design

AI-driven tools can significantly expedite the drug discovery process, from the initial screening of compound libraries to guided drug discovery to AI agents that can conduct drug screening (Figure 5). In the context of cardiovascular diseases, AI-guided drug discovery offers unprecedented opportunities to identify novel therapeutic targets, predict drug efficacy, and optimize the design of lead compounds.¹²⁴

AlphaFold, developed by DeepMind, has revolutionized protein structure prediction, which is critical in understanding drug-target interactions. AlphaFold and subsequent models like AlphaFold2 and ESMFold can predict protein structures, enabling a detailed map of the protein landscape, which in turn facilitates the identification of drug-binding sites and the design of more efficacious therapeutic molecules.^125,126 Integrating AlphaFold’s capabilities with drug discovery platforms has already led to identifying novel small molecule inhibitors for targets like CDK20, a kinase implicated in various biological processes, including those relevant to cardiovascular diseases.¹²⁷

Beyond predicting the structure of proteins, generative AI models are now being employed to design new proteins and peptides as novel therapies.¹²⁸ These models have started transforming the approach from a traditionally reactive process to a proactive one. By incorporating principles from diffusion models used in synthetic image generation, researchers have expanded the scope of generative AI to protein design. This involves developing models to iterate on representing the protein’s structure until a desired configuration is achieved, mirroring the process of natural protein folding but directed towards a specific functional goal.¹²⁹ Furthermore, the design of targeted drugs may be further optimized through Bayesian platforms that can further incorporate prior knowledge to accelerate drug discovery and clinical application.¹³⁰ Thus, AI has the potential to accelerate the pace of drug discovery, offering a wealth of opportunities for cardiovascular medicine. From AI-driven biological insights to the imaginative design of biological therapies, the future of drug discovery is poised for a significant leap forward.

5b. AI for precision therapeutics: the expanding role of large-scale genomics

With the decreasing cost and improved availability of sequencing technologies, the era of ‘petabyte-scale’ genomics is upon us.¹³¹ In cardiovascular care, this means optimizing treatment regimens for patients based on their unique genetic makeup that influences drug metabolism, efficacy, and safety. Even without AI, genomics has begun to play a role in routine clinical care. For instance, pharmacogenomics has been incorporated into cardiovascular medicine, influencing therapeutic strategies for hypercholesterolemia, acute coronary syndrome, and cardiomyopathy.^132,133 Mutations in genes like TTR, linked to hereditary cardiac amyloidosis, can guide the use of specific treatments only beneficial to patients with certain mutations, underscoring the need for genetic testing in clinical workflows.¹³⁴ However, with the advent of multimodal AI, the opportunities to transform care and optimize therapeutics using genomics are now numerous.

First, population-scale genomics can pave the way for simulating responses to treatments using high-fidelity digital twins that incorporate genomics and response to treatments noted in historical data. This approach could significantly enhance precision medicine with the advent of genomically-guided individualized care.¹³⁵ Polygenic risk scores, as well as their genomic inputs, which capture inherent risk, can now be used as elements within multi-modal risk prediction models and as influencers of therapeutic response.¹³⁶ The ability to leverage genomic data at scale has required ML to be computationally mature and models to be flexible enough to allow genotypic and phenotypic data to be modeled simultaneously. Finally, current cardiovascular foundation models only utilize clinical data, but incorporating genomic data could lead to an understanding of different axes and scales of biology.

5c. AI for clinical discovery: personalized inference from clinical trials

Beyond the role of AI in biological discovery and optimizing clinical operations, it can also enhance learning precision care from large-scale experiments in the clinical domain in the form of randomized clinical trials (RCTs). Whereas the traditional interpretation of an RCT relies on estimating an average treatment effect across the range of included individuals, conceptually, a patient’s characteristics may modify the efficacy of an intervention.¹³⁷ Since in most RCTs, individuals are only assigned to one arm, the counterfactual scenario of the same patient receiving a treatment and a control is not knowable.

Thus, an unmet need in our broader quest towards precision medicine is to develop data-driven ways to personalize our estimation of the expected harm and benefit of various interventions. Statistical ML approaches provide us with a set of rules that can estimate counterfactual observations, thus providing a robust way to assess treatment heterogeneity independent of human biases. In one example, computational embeddings (“phenomaps”) of clinical trial populations can be constructed to simulate a universe of potential patient phenotypes, as represented in the original clinical trials,^63,138,139 which may flag higher-order interactions not reflected by univariate analyses. Using these phenomaps, in silico simulations, and iterative analyses, one can then estimate a personalized treatment effect estimate for a new individual, further incorporating metrics of representativeness and uncertainty.¹⁴⁰ Alternative approaches also adapt ML algorithms, such as causal forests, allowing modeling interactions between treatment assignment and baseline phenotypes.¹⁴¹ There have already been demonstrations of their role in identifying treatment effect heterogeneity, such as in patient selection for intensive blood pressure lowering,^63,64 and the appropriate use of cardiovascular risk-reduction therapies for diabetes mellitus,¹³⁹ among others. Collectively, these techniques can guide the translation of RCT findings into real-world populations but also facilitate next-phase transitions in RCT design, for example, defining appropriate phase III inclusion criteria using phase II inference.

Beyond data collected from RCTs, causal ML approaches can be applied to observational data to estimate heterogeneous treatment effects.¹⁴² While such approaches only apply under strict assumptions, they could leverage the large variability already in clinical practice. These estimates from data already available could then be used to inform the design of prospective RCTs,¹⁴³ or enable the generation of evidence where RCTs may not be feasible due to time, cost, or safety-related considerations.

6a. Highly individualized targeted therapeutic selection and design

The advent of multiomics, encompassing genomics, proteomics, metabolomics, and other omics technologies, has been integral to the evolution of cardiovascular risk assessment and outcome prediction. AI can leverage these multi-dimensional datasets to uncover complex biomolecular interactions and pathways contributing to disease pathogenesis, prognosis, and response to therapy. A key application of AI in multiomics is the identification of biomarkers for cardiovascular risk, integrating multiomic data to delineate patient subpopulations with unique molecular signatures, paving the way for personalized medicine. These biomarkers can predict cardiovascular events, suggest individual responses to specific treatments, and define novel therapeutic targets.^144,145 Such AI-driven platforms not only identify novel drug targets but also provide a foundation for repositioning existing drugs to treat specific patient endotypes, potentially transforming the treatment landscape for cardiovascular and related metabolic conditions.¹⁴⁶ AI’s ability to distill multi-layered omics data into coherent patterns heralds a new era of predictive models that offer greater specificity and sensitivity for cardiovascular risk stratification and therapeutic intervention. This approach not only facilitates early intervention strategies but also refines our understanding of disease progression and therapeutic response at a granular level.

Integrating genomics, proteomics, and metabolomics with AI promises to reshape personalized drug design in cardiovascular care.¹⁴⁶ This would entail evaluating known treatments and potentially repurposed therapies and defining a need for custom therapeutic agents. For this, foundation models can be built linking these multiomics data streams with clinical outcomes. With these multiomic foundation models, personalized drug design hopefully becomes a reality.

Moreover, the implementation of such models facilitates the exploration of complex biomolecular networks that underlie cardiovascular conditions, identifying potential new drug targets that are specifically tailored to the nuanced needs of individual patients. This leads to a more efficient drug development process, reducing the time and cost typically associated with bringing new cardiovascular drugs to market.

6b. Acceleration of evidence translation through AI-powered clinical trials

A major barrier to translating discoveries into practice is the pace of conducting large-scale clinical trials to confirm their effectiveness and safety, with the typical development time for an innovative drug estimated at over 9 years.¹⁴⁷ Moreover, these trials typically provide evidence of benefit for 10 or fewer individuals per 100 participants treated, leaving the vast majority uncertain whether the benefit will be actualized. Acknowledging the need to improve the speed, efficiency, and representation of RCTs, AI tools are being developed to address key bottlenecks of this process.

First, AI is automating the eligibility screening process through ML- and NLP-based approaches that automate the mapping of patient characteristics to inclusion and exclusion criteria (i.e., Trial Pathfinder, Criteria2Query), allowing efficient screening of large repositories and EHRs.¹⁴⁸ Second, with multiple medical therapies competing for the same patient population, AI-guided platform trials may provide a data-driven inference mechanism to guide the allocation and matching of participants to individual strategies based on a priori knowledge regarding the interaction between their disease phenotype and anticipated treatment effects.¹⁴⁹ Third, advances in LLMs may streamline operations, accelerating and standardizing time-consuming and error-prone processes such as clinical event adjudication.^150,151

Fourth, AI may propel innovation in adaptive clinical trials and data-driven predictive enrichment, an important unmet need, as suggested by the FDA.¹⁵² For instance, data-driven ML approaches may be incorporated into RCT protocols to screen for treatment effect heterogeneity based on accumulating signals of benefit and harm. In the context of an adaptive design, this paradigm could maximize RCT efficiency by guiding data-driven predictive enrichment in a timely fashion.¹⁵³

Finally, digital twin technologies that simulate personalized health and disease conditions and projected trajectories by incorporating data with clinical knowledge may provide efficient alternatives to the somewhat limited traditional trial emulation methods from observational data.^67,154 These examples showcase how AI may soon provide end-to-end acceleration of all key steps of RCT design, enrollment, follow-up, and interpretation.

7a. How to evolve evidence-generation and regulation to evaluate personalization

The investment in precision care solutions requires a corresponding investment in strategies to test this personalization of care effectively. This is inherently challenging as these strategies are specific to the population they are tested in. By design, regardless of the strategy to define the precision care solution (genomic, metabolomic, and/or phenomic), the characteristics of the population will determine the range of phenotypes and the possible response to therapies. Therefore, prospective clinical studies and RCTs that deploy a personalization strategy in candidate individuals will require both broad coverage of phenotypes and a commitment to being evaluated in multiple distinct populations and care settings. This also includes defining appropriate continual learning mechanisms, particularly for sensor-derived data, that would enable their robust performance despite unavoidable data drifts. Moreover, since a majority of the data on these novel personalization strategies are derived from selective populations,¹⁵⁵ the omics profiles of key demographic groups vary widely,¹⁵⁶ and most of these studies are based on observational assessments,¹⁵⁷ rigorous prospective evaluation with a focus on clinically relevant outcomes is essential in different population groups. Finally, AI may also be used to confirm the validity of published data and information, thus providing an additional layer of scrutiny to evidence that may impact clinical practice.¹⁵⁸

However, given the resource intensiveness of conducting large-scale trials across care settings, especially with the increasing volume of algorithmic care, there will need to be a focus on nimble, pragmatic trials that leverage AI to enhance the efficiency of the process. Ultimately, the approach to personalization, the populations at risk, and how it affects their outcomes are critical to defining their range of use and enabling their eventual translation to clinical care. These decisions will also define the confidence in these precision care solutions and enable a degree of reliability critical for the regulatory pathways.

7b. How do we address health equity, bias, and barriers to implementation

While the central premise of integrating AI in clinical care is to make care delivery simpler, safer, more accessible, and of better quality at lower cost, there are major barriers to realizing this vision. A key barrier is that the development and deployment of technologies aimed at improving cardiovascular health still rely on many data streams only obtained in a subset of the population.^40,41,159 Specifically, the more affluent have better access to wearable devices,³⁹ and only those with access to regular care undergo diagnostic testing over which AI will be deployed.¹⁶⁰ Moreover, since AI-driven care enhancements are currently largely seen as optional tools that a practice may choose to take up but are not mandated or often even covered by insurance,¹⁶¹ those that cater to a more resource-rich population are likely to bring them forward. A big frontier for AI-driven care innovation is to reach people where they live, using technology deployed in a way that they receive the benefits and with care processes aligned with their goals. This has to achieve balance with the business models that support technology diffusion, especially as AI-powered technology becomes an integral part of care. While ensuring broad access to high-quality care, there will also be a need to build trust in communities that have historically mistrusted medical innovations both due to systemic malpractices against their communities as well as challenges with health and technology literacy.¹⁶²

Similarly, even in health settings where novel methods will define precision care solutions, combining information drawn from clinical trials, genomics, or multiomics, such an approach could exacerbate disparities by being largely accessed by those with additional resources or seeking care at select institutions that offer them. Therefore, before their selective use becomes the norm, clinical guidelines, payers, and health systems must coalesce around the appropriate standards to enable their broad and equitable use.