Artificial Intelligence in Cardiovascular Clinical Trials

Abstract

Randomized clinical trials are the gold standard for establishing the efficacy and safety of cardiovascular therapies. However, current pivotal trials are expensive, lengthy, and insufficiently diverse. Emerging artificial intelligence (AI) technologies can potentially automate and streamline clinical trial operations. This review describes opportunities to integrate AI throughout a trial’s life cycle, including designing the trial, identifying eligible patients, obtaining informed consent, ascertaining physiological and clinical event outcomes, interpreting imaging, and analyzing or disseminating the results. Nevertheless, AI poses risks, including generating inaccurate results, amplifying biases against underrepresented groups, and violating patient privacy. Medical journals and regulators are developing new frameworks to evaluate AI research tools and the data they generate. Given the high-stakes role of randomized trials in medical decision making, AI must be integrated carefully and transparently to protect the validity of trial results.

CENTRAL ILLUSTRATION

Opportunities to Improve Clinical Trials With Artificial Intelligence

Artificial intelligence (AI) research tools have the potential to improve clinical trials at multiple stages including design, recruitment, follow-up, and interpretation. EHR = electronic health record; NLP = natural language processing.

Randomized clinical trials are the gold standard for establishing the efficacy and safety of medical therapies.^1–3 Evidence from randomized trials supports the regulatory approval or clearance of novel drugs and medical devices, as well as the clinical practice guidelines and insurance coverage determinations that govern whether, when, and in whom such therapies are used. Unfortunately, pivotal randomized trials are expensive and lengthy, and they frequently fail to include a group of patients representative of those who will ultimately receive therapy.⁴ These limitations are particularly salient in cardiovascular (CV) medicine, where trials are designed to detect relatively modest reductions (eg, 20%) in outcomes that accrue slowly, such as myocardial infarctions, heart failure events, or CV deaths.

The cost of clinical trials supporting regulatory approval of new agents in CV medicine has been reported to be $35,000 per participant or more.^5,6 The several-year duration required for pivotal trials may delay potential benefits for patients and also reduces the time between regulatory approval and patent expiration, the window in which pharmaceutical companies must generate a return on research and development expenses that often exceed $1 billion.⁷ Despite the cost, marginalized racial and ethnic groups and women are underrepresented in clinical trials, a factor that undermines the equity, generalizability, fairness, and credibility of trial results.^8–11 Thus, clinical trials must evolve to be less expensive, faster, and more representative of diverse patient groups.

Artificial intelligence (AI) has affected many areas of society, spurred by technical advances in deep learning and large language models. The release of public-facing tools such as ChatGPT (OpenAI) demonstrated the utility of AI to individual users. In biologic science, the potential value of AI has been demonstrated by, for example, deep learning models that accurately predict protein structure and have rationally designed a new class of antibiotics.^12,13 AI has begun to be tested (although rarely applied) in clinical CV care, particularly for automated imaging interpretation and early disease diagnosis.^14–19 Although there is much promise, greater interpretability, validation, and monitoring of AI performance in a clinical environment are needed to support AI uptake.^20,21

AI has the potential to accelerate and automate clinical trials throughout their life cycle, from initial planning to patient recruitment, informed consent, ascertainment of endpoints, and dissemination of the results (Central Illustration).²² To date, AI tools have infrequently been applied within CV trials. Given the high stakes of clinical trials in evaluating new therapies, there is understandable concern that AI could introduce bias, reinforce existing inequities, or instill inaccuracy or inconsistency compared with traditional approaches. Regulators and journals increasingly receive submissions describing AI tools and AI-generated results, which require new evaluation frameworks.²³

On March 15, 2024, the Heart Failure Collaboratory convened a special focus meeting to discuss the role of AI in CV clinical trials and therapeutic development. Key stakeholders from academia, industry, medical journals, and the U.S. Food and Drug Administration (FDA) addressed opportunities for AI to improve trial design, conduct, and interpretation, as well as challenges and risks. This review summarizes the discussion from this meeting. It highlights ongoing work and future directions for applying AI in various aspects of clinical trial design and execution, with a focus on CV trials.

TRIAL DESIGN

AI may be able to assist investigators in trial design, including selection of inclusion criteria and pre-specified subgroups. In oncology, the Trial Pathfinder AI tool was developed to emulate trial results using electronic health record data and inverse probability weighting.²⁴ Applying this tool with various sets of inclusion criteria suggested that oncology trials could achieve similar treatment effect HRs with broader inclusion criteria that would in turn facilitate faster enrollment and greater generalizability of results. However, in silico trials are not a substitute for testing an intervention in real patients, in a prospective randomized fashion. Emulation of previously completed trials has not always found the same results as the randomized study.²⁵ Moreover, trial emulation from real-world data requires that patients are treated with the investigational therapeutic in routine clinical care, and therefore this approach is not possible for novel therapeutics that are not yet approved.

SCREENING POTENTIAL PARTICIPANTS AT SCALE

Recruiting participants is frequently the rate-limiting step in trial progress. Screening potential participants for complex eligibility criteria is time-consuming and often delegated to research assistants. Cohort identification using electronic health record data is common, as are alerts that notify investigators in real time of patients who potentially meet eligibility criteria. However, these methods are typically limited to the evaluation of discrete data elements and require timely data availability. An automated tool able to access free text, imaging, and laboratory data could improve accurate participant identification, with the ability to screen hundreds of thousands of patients in the electronic health record. Such a tool could uncover eligible subjects not initially considered by a human reviewer, and it could more quickly exclude those patients found to be ineligible from the cohort lists. Natural language processing (NLP) and generative AI may help to move beyond discrete data (ie, left ventricular ejection fraction, estimated glomerular filtration rate) to more subjective and nuanced trial criteria (symptomatic heart failure or NYHA functional class).

Several models have emerged for this task. Rules-based language models that are founded on specific words or phrases have successfully extracted inclusion and exclusion criteria data from unstructured notes on the basis of specific words or phrases.^26–30 More recently, the RECTIFIER (Retrieval-Augmented Generation–Enabled Clinical Trial Infrastructure for Inclusion Exclusion Review) tool was developed and tested in the COPILOT-HF (Co-Operative Program for Implementation of Optimal Therapy in Heart Failure) trial, a randomized trial for patients with symptomatic heart failure.³¹ RECTIFIER used Generative Pretrained Transformer Version 4 and Retrieval-Augmented Generation to assess 6 inclusion criteria and 17 exclusion criteria in potential participants’ electronic health record data. RECTIFIER was inexpensive ($0.10 per patient screened) and accurate (its eligibility assessment agreed with expert clinicians in 98%-100% of cases). As AI increasingly standardizes documentation, automated annotation tasks may become easier and more accurate. Eligibility assessment at scale may represent a relatively low-risk application of AI because the final decision for enrollment resides with an investigator who can prevent the enrollment of inappropriate participants.

OBTAINING INFORMED CONSENT

Current informed consent conversations between investigators and patients are often perfunctory and may bias the interaction toward enrollment. Long consent forms written in obtuse legal or scientific language do not help participants understand the planned research.^32,33 Generative large language models have been successful in reducing the complexity and reading time of surgical consent forms.³⁴ Interactive chatbots could play a future role in improving the efficiency and quality of informed consent for clinical trial participation. Unlike human investigators, chatbots have unlimited time to interact with participants and answer their questions, as well as the capacity to assess the participant’s understanding objectively and adjust language, readability, and mode of interaction to fit the participant’s needs. Moreover, AI-based informed consent may reduce participant burden by negating the need for in-person study visits during business hours.

The Pediatric Mendelian Genomic Research Center observational study implemented an optional chatbot-based consent process on participant smartphones that used a predetermined script.³⁵ Patients electing chatbot consent had shorter consent interactions, often completed the consent outside regular business hours, and scored as well as patients who received traditional in-person consent information on a quiz assessing their understanding of the study.³⁵ Participant satisfaction with the chatbot consent process was high (86%). In a study of consent for surgical care (not research), ChatGPT-generated text describing the risks and benefits of common surgical procedures was more readable, complete, and accurate than text written by surgeons.³⁶

The use of AI to obtain informed consent raises ethical concerns. Chatbots may behave in a coercive or biased manner. The human connection between investigator and participant plays a vital role in building trust and maintaining respect for the participant’s autonomy and dignity.³⁷ Given these concerns, AI should only augment rather than replace the investigator’s role for randomized clinical trials.

CLINICAL ENDPOINT ADJUDICATION

Automated adjudication of clinical outcomes by NLP has the potential to improve the trial cost, speed, and reproducibility. In current pivotal trials, outcomes such as heart failure hospitalization or myocardial infarction are commonly adjudicated by a central clinical events committee (CEC) of physicians who review participant medical records on the basis of established criteria or alternatively by site investigators.³⁸ CEC adjudication is labor-intensive, expensive, and not easily scalable, and individual site investigator decisions may not be uniform.

Early studies using NLP for clinical outcome adjudication focused on observational electronic health record data.^39–42 In the INVESTED (Influenza Vaccine to Effectively Stop Cardio Thoracic Events and Decompensated Heart Failure) trial, investigators externally validated an NLP model developed at a single center, Mass General Brigham (Boston, Massachusetts, USA), compared with human CEC adjudication. The NLP adjudication of heart failure agreed with the CEC in 87% of cases, thus demonstrating the model’s generalizability from the single center in a multicenter setting within the United States and Canada (Figure 1). Fine-tuning the model within INVESTED improved performance up to a reproducibility level equal to that of human reviewers.⁴³

FIGURE 1. NLP for Automated Endpoint Adjudication in the INVESTED Trial.

A natural language processing (NLP) model developed to identify heart failure (HF) hospitalizations in a single-center electronic health record (EHR) cohort was externally validated in the multicenter INVESTED (Influenza Vaccine to Effectively Stop Cardio Thoracic Events and Decompensated Heart Failure) clinical trial. CEC = clinical event committee.

AI has the potential to enhance outcome ascertainment in large-scale pragmatic trials in which human adjudication is not feasible. AI approaches are rapid, avoid delays in detecting treatment benefits or harms, and may provide more consistency than site-level adjudication.^44,45 Top priorities for research and development in outcomes ascertainment include developing more accurate models, assessing generalizability to international trials or CV endpoints beyond heart failure, and evaluating strategies that combine human and AI chart review.⁴⁶ Moreover, in addition to efficacy outcome adjudication, AI could categorize adverse events, ascertain relevant and possible underdiagnosed comorbidities, or identify patients who are at high risk for dropping out and who may benefit from additional education.

DIGITAL BIOMARKERS

Mobile health and wearable technologies now enable the collection of vast amounts of physiological data from trial participants without the burden of in-person specific study visits. Digital biomarkers can be obtained from such devices, including metrics derived from vital sign measurements, skin temperature, skin conductance, and accelerometry, among others (Figure 2).^47–49 Consumer wearables measure heart rate more consistently and precisely than in-clinic measurements, and they can capture diurnal variation, thereby offering even more information than in-clinic measurements.⁵⁰ Emerging large language models refined to analyze data from wearables, such as the Personal Health Large Language Model from Google, may help extract meaningful insights and recommendations for patients.⁵¹ Global positioning systems on smartphones can assist investigators in determining when a patient visits a clinic or hospital by using geofencing.⁵² The HearO speech analysis tool (Cordio) has been developed to identify congested compared with euvolemic patients with heart failure from recordings of their speech.^53,54 Noninvasive tools to estimate pulmonary congestion from handheld devices are also in development.⁵⁵ These tools may help identify worsening heart failure events in clinical trial participants remotely.

FIGURE 2. Applications of Digital Health Technologies for Remote Data Collection.

Digital health technologies may aid in the collection of physiological and patient-reported outcomes. GPS = global positioning system

Translating digital biomarkers into clinical practice or research applications requires rigorous verification and validation, which, in the most straightforward implementation, involve correlation with traditional surrogate markers. With AI, there is an opportunity to discover novel digital biomarkers that may not have a comparable traditional metric. Both the comparative and novel paths of digital biomarker discovery require establishing a link to outcomes that are meaningful to clinicians and patients.⁵⁶

Digital biomarker data collection often depends on patients’ owning and being comfortable using digital technology. This requirement may lead to exclusion of patients who are poor or who lack technological literacy, and it may thereby exacerbate the underrepresentation of some socioeconomic groups. One solution is to provide a parallel option for patients that does not require technology.

AI-ENABLED INTERPRETATION OF CARDIOVASCULAR IMAGING

CV imaging may be included in clinical trials as an inclusion criterion or outcome, or for safety monitoring. Central core laboratories often provide standardized reviews of imaging studies within trials, but these reviews are labor-intensive, costly, and time-consuming. Advances in deep learning applied to images have catalyzed the development of numerous models for automated interpretation of CV imaging studies.¹⁴

Within clinical trials, automated imaging interpretation can potentially reduce the cost and improve the reproducibility of imaging-based outcomes. For example, recent clinical trials of cardiac myosin inhibitors for hypertrophic cardiomyopathy required frequent echocardiograms to assess the drug’s effect on left ventricular systolic function and left ventricular outflow tract obstruction and determine dose adjustments.⁵⁷ Instantaneous interpretation of these echocardiograms by using validated AI tools could have eliminated the time required for the echocardiograms to be read, or even needed, and, therefore, could have shortened the trial’s duration and reduced costs. In echocardiography, deep learning models accurately measure left ventricular ejection fraction and other parameters.^58–61 Deep learning models for interpreting electrocardiograms,¹⁷ cardiac magnetic resonance,^62,63 and computed tomography⁶⁴ are available and could be similarly applied in clinical trials. AI interpretation of an electrocardiogram may be a proxy for the echocardiogram and may even obviate the need for an echocardiogram.^65,66 Deploying AI for image interpretation at the point of care is an important hurdle to meaningful patient impact. Open-source tools such as the AI-integrated Picture Archives Communication System (PACS-AI) platform can help integrate AI into existing imaging systems and thereby facilitate safe deployment.⁶⁷

Automated assessment of cardiac catheterization images could enable rapid screening of patients with coronary artery disease for trial eligibility. CathAI and DeepCoro are deep learning tools that accurately measure stenosis severity on coronary angiograms.^68,69 Several companies, such as HeartFlow and Cleerly, have algorithms that automate assessments of stenoses and flow. CathEF measures left ventricular ejection fraction from standard angiographic videos with a mean absolute error in ejection fraction of 7% to 9% compared with echocardiographic measurements simultaneously.⁷⁰ These tools, and many others that will be developed, will help overcome the challenges of recruiting and randomizing research patients in the catheterization laboratory immediately after initial angiography without delaying treatment.

CONTINUOUS PARTICIPANT MONITORING OUTSIDE INDIVIDUAL SITES

Explanatory clinical trials typically collect participant follow-up data (vital signs, laboratory testing, and patient-reported outcomes) from clinical electronic health records and intermittent in-person visits. For clinical event outcomes, medical records are printed, scanned, and manually submitted for adjudication by research core laboratories or committees, such as the CEC. The move to decentralize clinical trials seeks to shift the evaluation and monitoring of patients from the research site to the patient’s home or other convenient setting.^71–74 Decentralization can reduce the considerable site-related and socioeconomic burdens for patient participation, such as transportation, time off work, and child care. Decentralization expands the geographic reach of clinical trial participation to patients who do not live near academic medical centers. Lower participant burden may lead to faster and more diverse recruitment, including of groups traditionally underrepresented in clinical trials.

AI is expected to facilitate the transition of trials from in-person site-based study-specific research visits to in-home or other ambulatory assessments. For example, an AI-enabled chatbot that obtains high-quality consent by smartphone obviates the study visit where traditional consent would be signed. Continuous at-home vital sign monitoring by wearables could eliminate in-person postrandomization study visits.^49,50 Direct access to a participant’s electronic health record and automated endpoint adjudication eliminates the burden of working with the participant to identify clinical events and obtain clinical documents and is being integrated into trials.⁷⁵ AI can improve the data processing and incorporate a vast corpus of records to distill key information vital to the trial.⁷⁶ In addition to lowering costs and reducing participant burden, these methods generate a continuously updated and richer data set. Continuous and blinded evaluation of trials by AI algorithms may provide faster feedback to the sponsor and regulators for signals of potential treatment benefit or harm, facilitate adaptive trial design, and complement the current practice of episodic reviews by a human data safety monitoring board.

ANALYZING WHICH PATIENTS BENEFIT MOST FROM THERAPIES

Identifying subgroups of trial participants who may have benefited more from the investigational therapy is a key question in clinical trial analysis and is critical for personalized medicine. A traditional approach is to prespecify a handful of biologically plausible subgroups and test whether these baseline variables modify the effect of treatment. However, trials are rarely powered to test for such interactions.

AI has the potential to identify subgroups of patients who respond to therapy by using all pre-randomization features. One such methodology creates a multidimensional representation of the trial group across all baseline characteristics, identifies “neighborhoods” of similar participants, and quantifies a participant’s likely treatment response on the basis of the response of similar participants. This approach has demonstrated accuracy in predicting treatment responses in external validation data sets for aggressive blood pressure management trials, anatomic vs functional testing for coronary disease, and sodium-glucose cotransporter 2 inhibitors.^77–79 Such an analysis of expected treatment effect could be applied not only for post hoc interpretation and personalized medicine, but also during ongoing trials to set and adapt eligibility criteria to enrich for participants likely to benefit and to limit harm in patients less likely to benefit and thus more fully power analyses in the selected subgroup. Simulation analyses suggest an adaptive eligibility strategy could reduce the necessary sample size by 15% to 20%.⁸⁰

PUBLICATION AND DISSEMINATION OF RESULTS

Generative AI is already used to hasten the preparation of academic manuscripts. After the release of ChatGPT, the academic publishing community was forced to grapple with whether using generative AI to draft, edit, or review papers is ethical and safe.⁸¹ Many journals, including the New England Journal of Medicine and JACC, have coalesced in permitting generative AI in manuscript preparation as long as the authors disclose its use and take full responsibility for the final manuscript.^82,83 Science, which initially forbade any use of generative AI, relaxed its policy in line with this consensus.⁸⁴

Applying generative AI to automate additional post-trial activities from data analysis to preparation of regulatory submissions could meaningfully reduce the time from trial completion to availability of effective drugs to patients.²² Moreover, AI could assist with communicating trial results back to trial participants in nonmedical language, a responsibility that has often been neglected. Each of these steps presents, however, a potential risk for biased, fraudulent, or simply incorrect interpretation of the trial results. The gradual integration of AI with careful human oversight remains the most prudent path to improving efficiency while maintaining safety and refining best practices.

ROLE OF MEDICAL JOURNALS IN EVALUATING AI METHODOLOGY

Academic journals also have a key role to play in vetting and popularizing AI tools.^85,86 Studies assessing AI’s technical accuracy and clinical impact should be subjected to peer review. Papers on AI tools pose several new challenges for journal editors. First, because AI technology is progressing quickly, journals must offer prompt review and publication, given that the field is evolving rapidly. Second, journal editors must be open-minded to novel research methods that challenge the status quo and recognize the need to find qualified reviewers. Third, although some journals contend that authors must release underlying models publicly at publication, this may not be feasible when the software is proprietary intellectual property. An alternative to making the models public is to require validation by a qualified, independent party with full access to the software model and data sets. This is particularly important because developers of AI tools often have a financial interest in them; journals should continue to insist that all authors disclose financial ties.⁸⁵ Independent “health AI assurance laboratories,” federal agencies, or academic groups may be positioned to evaluate AI tools credibly.²⁰

REGULATORY PRIORITIES AND GUIDANCE ON AI IN CLINICAL TRIALS

Since 1995, the FDA has received >300 submissions for drugs and biologic products with AI components and >700 submissions for AI-enabled medical devices.^23,87,88 The drug and biologic application submissions using AI traverse the landscape of drug development from drug discovery to postmarket safety surveillance and cut across a range of therapeutic areas, including CV disease. The diverse uses of AI in these submissions highlight the need for careful regulatory assessment of both benefits and risks and underscore the importance of adopting a risk-based approach commensurate with the level of risk posed by the specific context of use. For any specific AI application in drug development, model risk calculations will be determined by model influence and the decision consequence on the basis of the context of use. For example, high-risk models may require more evidence of credibility than low-risk models, and the regulatory approach may differ accordingly.

As with any innovation, AI creates opportunities and new and unique challenges. To meet these challenges, the FDA has accelerated its efforts to create an agile regulatory ecosystem that can facilitate innovation and adoption while ensuring public safety and guarding against potential risks.⁸⁹ For example, in 2021, the FDA, together with Health Canada and the United Kingdom’s Medicines and Healthcare Products Regulatory Agency, jointly identified 10 guiding principles informing Good Machine Learning Practice for medical devices that are AI enabled.⁹⁰ These principles include the importance of bias mitigation by ensuring that clinical study participants and data sets are representative of the intended patient group using the device, the performance of the human AI team, and postapproval monitoring of performance. In March 2024, the FDA’s medical products centers published a joint report describing 4 major areas of focus in regulating AI (Table 1).⁸⁷ These areas are collaboration with key stakeholders, support for innovation, development of standards and best practices, and research related to monitoring AI performance for bias and inequity.

TABLE 1.

U.S. Food and Drug Administration Areas of Focus Regarding the Development and Use of AI Across the Medical Product Lifecycle

1. Foster collaboration to safeguard public health.

2. Advance the development of regulatory approaches that support innovation.

3. Promote the development of harmonized standards, guidelines, best practices, and tools.

4. Support research related to the evaluation and monitoring of AI performance.

Reprinted from the U.S. Food and Drug Administration.⁸⁷

AI = artificial intelligence.

The regulatory process for AI tools targeted to research rather than patient care, or a combination of the 2, is distinct from the 510(k), de novo, or premarket approval medical device review but follows similar principles. The FDA Drug Development Tool (DDT) and Medical Device Development Tool (MDDT) programs specifically assess AI tools for use in research rather than clinical care. The DDT program from the Center for Drug Evaluation and Research (CDER) and the Center for Biological Evaluation and Research (CBER) qualifies methods, materials, and measures that have the potential to facilitate drug development.⁹¹ Examples of non-AI DDTs include biomarkers for clinical trial enrichment, clinical outcome assessments to evaluate clinical benefit, and animal models. DDT qualification is not required for applying a tool in clinical trials. Nonetheless, it may streamline the review of subsequent regulatory submissions by avoiding the need for the FDA to reconsider and reconfirm the tool’s validity in each drug program to which it is applied. DDTs are evaluated for a specific context of use, including recognition of a tool’s limitations and the contexts in which its application is inappropriate. The Innovative Science and Technology Approaches for New Drugs (ISTAND) Pilot Program accepts submissions for DDTs focused on AI and digital health technology (DHT).⁹² The MDDT program at the Center for Devices and Radiological Health (CDRH) is similar to the DDT program but focuses on medical device research and regulation.⁹³ These programs do not replace the traditional dialogue between investigators or sponsors and regulators to review clinical trial methodology in advance.

Additional FDA programs provide opportunities for early engagement with the agency regarding applying AI tools in clinical trials. For example, the Critical Path Innovation Meetings (CPIM) facilitate dialogue among industry, academia, patients, and government regarding new tools and technologies to improve the efficiency of drug development. These nonbinding discussions seek to encourage innovations, including potential submissions to the DDT program. Another example of a pathway for engagement is the DHT Program, which enables developers of AI-enabled DHT used in a drug development program (eg, wearables that collect data remotely and therefore reduce the burden on trial participants) to interact with the agency.⁹⁴ Further dialogue on approaches to qualifying such technologies is planned. The recently formed CDER Center for Clinical Trial Innovation seeks to improve the efficiency and effectiveness of clinical trials, participant diversity, and the pace of drug development, and it may provide an additional forum to discuss best practices for integrating AI. In these programs, the FDA emphasizes the importance of early discussions between stakeholders and the agency regarding using novel AI technologies in clinical trials.

LIMITATIONS AND POTENTIAL PITFALLS OF AI IN CV TRIALS

Despite the tremendous promise of AI in improving CV trials, several potential risks must be acknowledged and managed carefully (Table 2). First, AI may be less accurate than time-tested data collection and interpretation methods by humans. Given the importance of clinical trial results to patient care, using “quick and dirty” methods to save cost or time carries even greater risk. Second, AI is susceptible to data set shift, in which differences between the data used to train and validate the model and the data on which the model is applied lead to poor performance.⁹⁵ For example, an event adjudication model developed in one electronic health record system may perform poorly after a site transition to a new system. Stable models may become quickly outdated, and continuously learning models may accumulate new biases or experience degraded performance. Approved AI technologies require a plan for monitoring accuracy and bias at a routine cadence for as long as they are in use.⁹⁶ Third, AI tools that learn from completed trials may encode or amplify biases against women or underserved groups.⁹⁷ For example, a model seeking to identify eligible participants that was trained on trials with very low representation of Black patients may systematically ignore eligible Black patients and thus perpetuate inequity. Mitigating bias requires ensuring that the data used to train and evaluate the AI tool are representative of the diverse groups of patients to which it will ultimately be applied. Large, publicly available data sets such as CheXpert for chest radiographs,⁹⁸ MIMIC-IV for electronic health records,⁹⁹ and EchoNet-Dynamic for echocardiograms,⁵⁸ can help mitigate bias and promote transparency. Evaluation for and mitigation of bias must occur through the life cycle of the tool, including the postmarket phase.¹⁰⁰ Fourth, the confidentiality of participant medical data must be protected. For example, generative AI models automatically use all submitted data for future training unless specific restrictions or safeguards are in place, such as limitations on the basis of patient privacy or consent. Emerging open-source large language models that can be hosted locally within health care institution servers and tools such as LM Studio may help overcome privacy concerns.¹⁰¹ Fifth, human trial researchers relying on automated AI tools may not learn core competencies such as obtaining informed consent or adjudicating events, thus leaving the trial vulnerable if AI becomes unavailable or is flawed. For competencies critical to participant safety or the integrity of results, trials must ensure appropriate human oversight and competency.

TABLE 2.

Mitigating Key Risks of AI in Clinical Trials

	Risk	Mitigation Strategy
Poor generalizability	AI model is inaccurate when applied in a novel context of a prospective trial.	Validate AI models in the context in which they will be used.
Data set shift	AI model performance erodes with changes in medical practice or data organization.	Update the AI model with contemporary data guided by a predetermined change control plan.
Algorithmic bias	AI models may perpetuate or amplify biases against marginalized groups.	Train or validate the AI model on representative data, and evaluate for bias.
Lack of clinical interpretability	Digital biomarkers may be unproven surrogates for clinical outcomes.	Insist on proven relationship between biomarkers and clinical outcomes.
Patient data privacy	Sharing patient data with third-party AI providers risks breach of confidentiality.	Require strict privacy agreements and encryption.
Patient access to technology	Requiring participants to use digital technology may exacerbate biases in enrollment.	Provide backup options by which patients can enroll and be followed up without digital technology.
Loss of human competency	Future researchers relying on AI may not learn critical skills.	Maintain human oversight of tasks required for participant safety and integrity of results.

AI = artificial intelligence.

CONCLUSIONS

Randomized clinical trials of sufficient size are necessary to evaluate novel CV therapeutics and inform regulatory judgments, but traditionally they have been costly, long, and insufficiently diverse. Emerging AI technologies have the potential to address these limitations by automating and streamlining clinical trial operations. Opportunities to integrate AI exist throughout a trial’s life cycle, including identifying and obtaining consent from eligible patients, ascertaining physiological and clinical event outcomes, and analyzing or disseminating the results. Medical journals and regulators are developing new frameworks to evaluate AI research tools and the data they generate. Given the high-stakes role of randomized trials in medical decision making, AI methods must be integrated cautiously and thoughtfully to protect the validity of trial results.

ABBREVIATIONS AND ACRONYMS

AI

artificial intelligence

CDER

Center for Drug Evaluation and Research

CEC

clinical events committee

CV

cardiovascular

DDT

Drug Development Tool

DHT

Digital Health Technology

FDA

U.S. Food and Drug Administration

MDDT

Medical Device Development Tool

NLP

natural language processing