The Introduction of AI Into Decentralized Clinical Trials

Corresponding Author

Cardiovascular disease remains the leading cause of mortality worldwide, underscoring the critical need for ongoing advancements in treatment strategies. Clinical trials serve as the conduit for translating medical advancements into patient care, yet traditional trial methodologies are often constrained by size, duration, and resource limitations. Leveraging recent developments in remote monitoring, computational technology, and artificial intelligence (AI) presents a compelling opportunity to revolutionize cardiovascular clinical trials. This article proposes a framework to facilitate the adoption of novel approaches aimed at enhancing the accessibility, efficiency, and inclusivity of cardiovascular trials, ultimately accelerating the pace of therapeutic discovery and improving patient outcomes.

Home-based care, including decentralized clinical trials, allows for clinicians and researchers to meet patients where they are, improve recruitment of diverse populations, and reduce burden to patients and health systems.^1,2 With the paradigm shift of many components of clinical trials to home settings, accelerated by necessity during the COVID-19 pandemic, there is also an increase in the opportunity to engage and retain patient participation in clinical trials from the comfort of home. This may improve adherence and retention, especially for those who have geographic and logistical barriers to participation. Decentralized clinical trials provide the opportunity to reach traditionally underserved and underrepresented populations and geographies. Additionally, there is significant value derived by shifting to decentralized clinical trials methods, including shorter cycle time, fewer screen failure, and fewer protocol amendments resulting in significant return on investment required to adopt and train in the novel methodology.^1,3

Necessary components of decentralized care and decentralized clinical trials include remote patient monitoring through connected devices, asynchronous data collection, and electronic patient-reported outcomes collection. However, this increased data generation can create a “data deluge” and overwhelm sophisticated data analytics, statistical, and clinical teams. In order to meet the demand of further decentralization of care and clinical trials, the cardiovascular field will require additional evolution to integrate AI tools (Figure 1).

Figure 1.

Artificial Intelligence Opportunities in Decentralized Clinical Trials

AI holds promise at various stages of the trial process including recruitment, patient monitoring and retention, as well as post-trial monitoring and dissemination. The above depicts specific methods that may be used at each stage.

Cardiovascular trials are underrepresented in the use of AI, even though cardiology has a long history in remote monitoring and adoption of innovative technologies and methodologies. A recent meta-analysis of AI application in clinical trials showed that AI was most often applied in oncology trials, with far fewer cardiovascular trials utilizing AI techniques. AI tools create efficiencies, including reduced trial size and shorter duration, and of particular interest to decentralized trials, AI-based sensors and wearable devices improved patient monitoring. With increased data generation from remote monitors and electronic patient-reported outcomes collection, AI analysis facilitates the statistical needs of clinical trials.

To address the need for a risk-based regulatory framework that promotes innovation and protects patient safety, the Food and Drug Administration published guidance on usage of AI and machine learning (ML) in drug development. Per the Food and Drug Administration definition, AI is a branch of computer science, statistics, and engineering that uses algorithms or models to perform tasks and exhibit behaviors such as learning, making decisions, and making predictions. ML is a subset of AI that allows models to be developed through analysis of data, without models being specifically programmed.⁴ Across just about every industry, the last year has demonstrated that the field of AI is growing at an exponential rate.

On average, it takes 10 to 15 years and over $1.5 billion USD to bring a new drug to market, with over 50% of time and investment going toward the clinical trial phase. It is estimated that about 86% of all trials do not meet enrollment timelines, which take up to one-third of the trial duration. Furthermore, phase III trials which require the most patients have a 32% failure rate due to poor patient recruitment.⁵ Furthermore, the number of patients needed to establish a treatment effect is the top cost driver.⁶ Utilization of AI-based algorithms to optimize patient selection and recruitment shows promise in improving trial efficiency. These include population and adaptive enrichment, clinical trial matching, and targeted patient outreach.

ML methods for anticipating treatment effects based on the phenotypic diversity of patients and interventions in randomized controlled trials are being actively explored. The method involves creating a multidimensional representation of a population across prerandomization features, or “phenomap.” It then extracts signatures that define consistent risks or benefits in each trial arm. This method was validated in several trials.⁶ This method is promising because of the ability to enrich a cohort with patients who are most likely to respond to treatment.

Furthermore, ML was also evaluated as an application to adaptive trials that use phenomapping-derived study arms to predictively enrich participants.⁶ This allows the accumulation of data to change aspects of the trial without undermining its validity and integrity. This was recently performed as a post hoc simulation of real randomized controlled trial data from two major trials: 1) pioglitazone, IRIS [Insulin Resistance Intervention after Stroke trial; and 2) disease management strategy (intensive versus standard blood pressure reduction in the SPRINT [Systolic Blood Pressure Intervention Trial.⁶ They demonstrated that their approach adaptively enriched patients most likely to benefit from the intervention and reduced sample size up to 18% while preserving efficacy and safety signal.³ Enrichment strategies such as the above have also been described in trials involving heart failure with preserved ejection fraction.⁷

Clinical trial matching systems utilizing AI are also being developed. Natural language processing (NLP) techniques may be applied to the electronic medical record to synthesize structured and unstructured data and surface them to clinical decision-makers. ML and deep reinforcement learning (RL) can adapt and integrate feedback based on the quality of the output. Techniques such as these are being explored for analysis of electronic medical record and clinical eligibility databases for purposes of recruitment.⁵ Furthermore, use of NLP holds promise in a more efficient method of electronic health record data extraction as opposed to traditional manual chart review. A recent study examined an NLP model, the Community Care Cohort Project (C3PO) NLP model for heart failure, and compared it to the gold standard clinical events committee in an external multicenter clinical trial. The tailored C3PO and de novo NLP models demonstrated agreement of 93% (95% CI: 92%-94%) and κ of 0.82 (95% CI: 0.77-0.86) and 0.83 (95% CI: 0.79-0.87), respectively, vs the clinical events committee. Clinical events committee reviewer interrater reproducibility was 94% (95% CI: 93%-95%; κ of 0.85 [95% CI: 0.80-0.89]).⁸

Beyond the recruitment process, the shift toward digitalization of DCTs has also opened the door to AI opportunities to improve care delivery and retention. Trials average about 30% dropout rates which result in additional recruitment needs, delays, and additional cost.⁵ For patients engaged in a clinical trial, there is a growing burden of data collection requiring them to fill out surveys, diaries, and electronic clinical outcome assessments frequently. These are time-consuming, result in incorrect data input, can reduce adherence, and even lead to drop out.⁹ In fact, these overwhelming tasks lead to an average of 40% of patients being nonadherent after 150 days.⁵

Wearables and AI techniques can potentially be used to continuously collect data, relieving patients of data collection burden. Deep learning models can be used to analyze data and synchronously log relevant events, generating disease diaries. Deep learning models can intermittently be retrained with updated measurements and even become patient specific based on patient behavior.⁵

It has also been proposed that leveraging continuous data could be used for RL to understand trial participants’ engagement patterns, accommodate to those habits, anticipate, and prevent dropout. RL is a ML technique that trains software through a trial-and-error learning process to achieve optimal results. Similar techniques have been used to create chess playing computers to beat the best human opponents and by Meta to optimize Facebook notifications. It allows for customization of messaging toward the individual participant, reducing low yield notifications.⁹ This technique is also being explored to limit drug dosage to the smallest dose needed to achieve results without causing side effects specifically with chemotoxic drugs in the oncology space.⁵

There are numerous challenges that we will be forced to navigate as the field of AI in health care grows, particularly in clinical trials. Incentives for efficiency and cost-saving opportunities are likely to push this field exponentially in the coming years. Given the immense problem already faced in the literature as it relates to bias and inequitable representation, AI as applied to trial design must be met with caution and oversight without which disparities may be amplified. The following are some of the major challenges and ethical considerations as the field evolves:

• •Perpetuation of health inequities: Algorithms are dependent on the initial data sets used for training and may exaggerate bias if samples do not include diverse and truly representative data.
• •Lack of accountability: Without strict clinical oversight, leaning too heavily and too soon on AI may lead to lack of informed consent, coercion, or mistreatment.¹⁰
• •Privacy concerns: AI databases rely on large amounts of sensitive patient information making them vulnerable to misuse or breaches.

While decentralized care and AI offer promising avenues for enhancing clinical trials, their implementation in cardiovascular research demands robust evidence generation and use-case development. We advocate for increased adoption of these innovative approaches to promote diversity within trials. However, it is crucial for clinicians and researchers to remain vigilant regarding potential biases and limitations in the data used to train AI models. Cardiovascular trials have a unique opportunity to spearhead the integration of AI, enabling broader therapy access and improved representation.

Funding support and author disclosures

There are no relevant funding disclosures. Dr Goldberg is employed by Heartbeat Health. Dr Amin is employed by Bristol Myers Squibb and has an academic appointment at Weill Cornell Medicine, New York Presbyterian. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.