Three Ways AI is Revolutionizing Clinical Development

Image Credit: © Egor - stock.adobe.com

Artificial intelligence (AI) is the biggest transformation in life sciences in the last 25 years—an advancement akin to mapping the human genome, with potential for a much higher rate of scientific breakthroughs over the next decade. But there’s still a lot to consider when applying AI to clinical development and product commercialization, especially when patient safety must always be at the center of our thinking. Navigating the hype and the buzzwords associated with AI can be challenging for clinical development organizations.

For the past two decades, Phesi has been applying data science including AI, machine learning, and predictive modeling to clinical trials. From this experience, we’ve identified three use cases where AI can have the greatest impact on clinical development:

1. Patient selection and protocol optimization

AI enables advanced analysis of historical trial data and real-world patient data to create a digital patient profile (DPP), a statistical view of the patient attributes for the indication under study—including demographics, comorbidities, concomitant medications, and various disease related measures. The DPP is the starting block for any study. Conducting further analysis from the foundation of a DPP and related data increases the precision of research and findings, and ensures AI is patient-centric.

For example, the DPP can be used to identify a set of relevant protocols to highlight the ideal number of outcome measures and endpoints best suited for an indication and the specific target patient population. This level of insight is critical. A recent AI-based analysis of 2,401 Phase III trials found that studies often over-collect data through having too many outcome measures in a protocol, and that an optimized number of outcome measures (and therefore a lower burden on the patient, investigator sites and the sponsor) leads to better investigator site enrollment and shorter trial enrollment cycle time. Directly comparing type 2 diabetes trials from two different sponsors found that the sponsor using a median of 25 outcomes measures had lower investigator site enrollment performance and a longer enrollment cycle time compared to the sponsor with a median of only 10 outcome measures.

Clinical development organizations can also use DPPs to predict which types of patients are most likely to benefit from the trial and then target only those patients in the protocol. This identification of ‘responder’ vs ‘non-responder’ patients lowers dropout rates by avoiding the recruitment of patients who might become overburdened by the study, or will not respond to the treatment because of factors such as age, sex, or comorbidities. This also helps to improve diversity in clinical trials, by ensuring that sponsors are not always drawing from the same pool of patients, and are in fact recruiting participants from the patient subpopulations who best fit the actual demographics most affected by a disease.

2. Investigator site selection

Understanding the previous performance of investigator sites is critical for designing and executing a successful trial. Often, investigator site selection is based more on existing relationships and gut feel than precise, patient-led data analysis. For example, a Phesi analysis of 471 non-small cell lung cancer trials found that a fifth (20%) of the investigators running these trials specialized in different areas of oncology, with no history of recruitment in lung cancer studies.

While it is possible to use basic data analysis to determine a simple hierarchy of “most used” or highest-ranked investigator sites, that does not mean those sites are best-suited to a particular trial. Moreover, if all sponsors simply select investigator sites from the first tier of most used sites, those investigators will swiftly become overburdened, with a negative impact on enrollment performance.

Using lung cancer clinical trials in the US as an example, the median number of recruiting trials in a top 100 cancer hospital is 56, and the median number of recruiting trials in the top 100 investigators is 33. As a result, one in five investigator sites only enroll a single patient—prolonging cycle times, reducing return on investment (ROI) and putting data quality at risk.

Advanced AI algorithms deliver a more nuanced approach to selecting investigator sites. They can search through reams of historic and existing trial data to analyze factors like an investigator site’s clinical trial capacity, competing clinical trials, disease related clinical expertise (including treatment modality and interventions).

Combined with insight from a DPP, AI can identify investigator sites with a proven track record of accessing the specific patient population defined in a protocol. This adds another significant layer of precision in investigator site identification and country allocation, and can be quantitatively measured using a ‘patient access score.' For example, in acute ischemic stroke, an investigator site with a successful track record of conducting stroke clinical trials is important, but investigator sites that accessed acute ischemic stroke patients within a 48-hour time window from onset, as defined in a recent trial protocol, is critical for the success of the trial.

3. External control arms

External control arms, as an integral part of clinical development, present many potential opportunities for sponsors and patients alike—not least when it comes to the ethics of administering a placebo. AI-driven ‘digital twins’ have emerged as an innovative solution, allowing sponsors to replace a control arm with digital patient data.

AI can be used to build digital twins using the DPP for an indication as the basis of the analysis. The alignment between a protocol and the corresponding patient population derived from a DPP enables the generation of a digital twin that represents the patients recruited when following the protocol design at baseline. Safety and efficacy outcomes of the control arm as defined by the protocol can be analyzed, completing the construction of the digital twin.

Recent research has demonstrated the feasibility of this AI-led approach. A study published in Bone Marrow Transplantation in graft-versus-host disease, comparing a trial conducted with digital twins to existing standard-of-care treatment, hypothesized that results from the digital twin are statistically identical to the results from an actual clinical trial control arm.

Researchers’ first question when it comes to digital twins is often how likely they are to be approved by regulatory bodies. This is another area where showcasing a robust data foundation, such as a DPP, is critical. DPPs provide a structured adoption pathway for regulatory acceptance of digital twins and allow early, consistent engagement with regulators​. Further, the value of a digital twin goes beyond external control arms. Using a digital twin to better interpret results from a single arm clinical trial or using it as a historical control arm in regulatory submissions are among examples of other use cases.

The triangle of successful AI-driven clinical data science

Each of these approaches is individually powerful and serves as a tangible and achievable use case for companies seeking to adopt AI in clinical development. The next level of maturity—and where the true value of AI is unleashed for clinical development—comes with combining these use cases together.

The application of AI should be thought of as an integrated continuum, built from discrete use cases. When optimized protocols, effective country and investigator site selection, and digital twins are coordinated through the use of AI-driven clinical data sciences, clinical programs and trials are safer, faster, less burdensome and less costly (Figure 1). AI-driven clinical data science presents significant opportunity to the clinical development industry for lower attrition rates, faster cycle times, improved regulatory submissions, and a better foundation for future trials.

Figure 1. The triangle of AI-driven clinical data science; an integrated continuum of AI uses cases that accelerates clinical development and improves the success of outcomes.

Source: Phesi

The data foundation

One key takeaway is that all AI-driven clinical data science requires a foundation of verified and trusted data to be able to deliver benefits. Any data used to train clinical AI models should be:

• Comprehensive, covering as many investigator sites, patients, diseases and historical trials as possible—and at a large enough scale to be statistically significant.
• Contextualized, with an understanding of who collected the data, where and why, and ideally drawn from real-world sources.
• Verified, taken from credible sources, databases and registries.
• Complete, with full data parameters and containing fully reported results.

Finally, AI should be seen as a tool to increase efficiency, rather than a replacement for human expertise. AI supports the limitations of the human brain in terms of how much data we can process in one go, and can automate labor-intensive and time-consuming tasks to free capacity for researchers. But human experts are still needed to understand current medical practices and standard of care, to verify, contextualize, and decide how to apply results—and sometimes even push back on AI’s suggestions. In essence, the benefits of AI come not from the technology itself, but from applying it well.

AI has existed in some form in clinical development for decades. Now though, we are at a stage where algorithms have been refined to be more sophisticated than ever, unlocking new ways for technology and human expertise to work in concert and greatly improving ROI of clinical development operations. Though AI isn’t a magical solution to all ills, there are already many use cases that prove how powerful it can be for accelerating the development of impactful treatments for unmet medical needs. It promises to fundamentally transform today’s development and commercialization model.

Gen Li, PhD, MBA, is CEO and Founder, Phesi