Syneos Health

Michael Shim, National Medical Director- Fosun Pharma

Overview

There are numerous applications for artificial intelligence (AI) and machine learning (ML) in the life sciences industry, however not all of them are immediately profitable or relevant to an organization. How does a company go about finding the most productive use cases? We present a practical use case in our Medical Affairs business unit at Syneos Health® to help answer this question using current ML strategies.

One of our large biotech customers was behind on trial site activation for their current Phase 3 study. Our small team of one medical science liaison (MSL) and three Research Science Liaisons (RSLs) were tasked with delivering 35 activated trial sites within a short timeframe of four months.

Our team leveraged an unsupervised ML approach to perform a physician segmentation around this disease state. This data-driven approach allowed our teams to efficiently engage the right physicians who were most likely to participate in our customer’s trial. Our joint MSL-RSL team slightly over-delivered on our customer promise. At the end we identified and activated 36 clinical trial sites within the compressed timeframe.

Our Innovative Approach

Our ML approach contrasts with what has historically been done via a manual process. We compiled an initial list of relevant physicians using our licensed physician mapping platform. We were able to identify over 4,000 physicians. Thereafter, our team manually trimmed the list down to approximately 2,000 physicians based on past or existing contacts. We had insufficient information on the remaining 2,000 physicians, thus prompting us to explore a more focused physician identification strategy.

Our team used an unsupervised ML strategy to perform physician segmentation based on two variables. Before we move on, it is important to discuss what unsupervised ML is as it relates to our specific use case.

What Is Unsupervised Machine Learning?

Unsupervised ML uses ML algorithms to analyze and cluster unlabeled datasets. These algorithms discover hidden patterns or data groupings without the need for human intervention (IBM, 2024). Unsupervised learning (UL) has the ability to discover similarities and differences in data sets making it ideal for exploratory data analysis and in our case, physician segmentation.

Physician Segmentation Via Unsupervised Machine Learning

Our team took the initial list of 2,000 candidate trial sites and vetted each physician for two specific sources of information that would provide insight into their likelihood of participating in our customer’s study. The information we extracted include:

(1) The number of clinical trials in this disease state the physician had open at their practice.

(2) The number of relevant disease state publications a physician published over the past five years. We pursued this specific data-driven approach for the following reasons:

(1) Higher physician trial involvement could equate to a higher risk of competing studies.

(2) Physicians experienced in clinical trials could help ensure they could deliver on patient recruitment goals.

(3) Number of relevant disease state publications would correlate to relevant disease state expertise.

Targeting a cluster of physicians who had a modest number of relevant disease state publications with a minimal amount of actively enrolling trials appeared to be a rational and strategic approach. We did not rely on “luck” to find the trial sites.

Below we share a simplified model of our ML approach. It is not important to understand the ML programming in the following images. The main takeaway is that current ML technologies can be applied in practical ways to derive meaningful value for our clients.

Our 3 Step Machine Learning Approach

For simplicity, we demonstrate a representative sample of n=100 physician contacts. However, this same ML approach was used to assess all our candidate trial sites. The benefit of this approach is that you begin to see patterns and structure within the data which then translates to value. We utilized our Microsoft® Azure technology stack and our internal integrated development environment (IDE) to execute our ML algorithm.

RESULTS

Figure 4 shows how we parsed out the (0, 0) x-y axis physician segments as well as the outlying clusters. These physician segments were de-prioritized for the reasons we mentioned earlier. In our specific use case, this unsupervised ML approach provided us a physician segmentation that yielded approximately 800 priority physician targets. After four months of targeted field engagement our joint MSL-RSL team slightly overdelivered 36 new trial sites for our customer.

CONCLUSIONS Our ML use case demonstrates how a data-driven approach can bring tremendous value to an organization. Our ML solution helped our customer successfully get back on track with their clinical development timelines.

Our approach also helped to reduce cognitive burden for our team, thus ensuring efficient use of time and resources in other complementary valued-driven strategies for our client. Furthermore, the practical use of novel AI technologies empowered our small team with renewed confidence that ultimately helped in our successful business execution. This approach allows smaller MSL teams that have limited FTEs and resources to make a significant contribution and impact to a clinical development program.

As AI and ML systems continue to grow and become integrated across organizations it is imperative to use these technologies in a responsible and ethical way. Our internal AI Council provided legal and regulatory guidance on the use of our AI capabilities and output to help ensure necessary protections were in place for our customer and our organization.

Our ML approach can be scaled-out and utilized across different business units to complement and enhance traditional best practices. Implementing current ML innovations helped us optimize clinical trial site identification, which ultimately benefitted the patients we serve by minimizing delays to the normal clinical development process of new drugs.