At AbbVie, Machine Learning Aids Selection Of Clinical Trial Sites
By Deborah Borfitz
February 1, 2023 | At AbbVie, the introduction of machine learning into the site selection process for clinical studies is reaping tangible returns by increasing the number of sites meeting enrollment and startup time goals by more than 50%, according to Bardia Akbari, vice president of development operations in the oncology division. The algorithm helps study teams identify the best sites for a particular trial based on a quantitative analysis of their performance in similar studies done in the past.
Machine learning won’t solve every problem encountered when evaluating the feasibility of a study and predicting recruitment success, but it deserves to be in the mix, as Akbari will make clear during his presentation on the topic at the upcoming Summit for Clinical Ops Executives (SCOPE) in Orlando, Florida. This is particularly true when companies are entering a new therapeutic area or indication and qualitative information on the investigator pool is harder to come by, he notes.
The pharma industry has been using analytics to improve site selection for the past five to seven years and machine learning is one of the more recent additions to the arsenal of companies’ in-house IT and data scientist teams. The question is no longer whether to apply the concept to clinical operations but where and when to do so, Akbari says.
For more than a year now, machine learning has been a systemized part of the site selection process at AbbVie, motivated by inconsistencies in the number and performance of sites being tapped for trials, continues Akbari.
Moreover, the company’s “rather diverse” portfolio demanded a mechanism for coalescing available information on investigators and sites in one location to identify factors contributing to their success. The idea, he says, was to give individual study teams an overarching view of the options across all indications and studies.
In the past, site selection at AbbVie relied heavily on the expertise and experience of clinical research teams and medical groups and the expectation that good performance in the past was indicative of good performance in the future, says Akbari. Country-specific knowledge would be gleaned from local site monitors and medical liaisons who, based on their relationship with investigators on the ground, could provide information on a site’s patient population and observed capabilities.
It was widely acknowledged that a more efficient and objectively measurable process for selecting sites was required as the company forged deeper into its core therapeutic areas (oncology, immunology, eye care, virology, neuroscience, and aesthetics) and specialty disease areas (e.g., women’s health and cystic fibrosis)—and that machine learning was a viable means to that end, he continues. The analysis considers many of the factors that are important to physicians when predicting site performance.
From the standpoint of investigators, a big part of their decision to accept a study protocol in the first place is whether it offers a treatment option for their current patient base, Akbari says. If so, they additionally want to know that the investigational compound is novel enough and has a suitable safety profile given the disease state.
Furthermore, physicians need to believe that the comparator or the control arm is “good enough” for their patients in the event they aren’t randomized to receive the novel drug, says Akbari. They also need to know the study will not be too difficult to conduct in terms of the number of required procedures and visits for participants.
AbbVie’s multidimensional approach to site selection starts with the protocol concept sheet, which it terms a protocol synopsis since it closely resembles the final protocol, Akbari says. The synopsis provides sufficient information to establish whether investigators in different regions of the world would find the study attractive enough to enroll their patients in, based in part on its preliminary design and the scientific and clinical data generated to date.
Assuming there is sufficient presumed interest in a study in these pre-identified geographies, the draft protocol next gets more comprehensively evaluated and used for the selection of countries and sites. The process involves decision-making about what to prioritize—for example, how fast a site can be onboarded or if the study needs to be done in an academic center or community setting—creating a “building block of information that helps you select better sites as you go forward,” he says.
Changes to the protocol to accelerate enrollment or site performance are made on the front end, so as not to lose unnecessary months of valuable time in the middle of a study, continues Akbari. Prior to site selection, AbbVie has already established that the study is feasible based on health authority input and its traditional, intelligence-gathering work.
Real-world “analog” data is then used to match the current research scenario to similar clinical trials conducted in the past. Some of the earlier learnings are effectively applied forward, he explains.
The entire process, from initial assessment to site selection, generally takes between 10 and 12 weeks, Akbari reports. Once data is curated for the machine learning team, site identification and actual selection happens seamlessly in a short period of time. The company’s business insight team does the analytics work in collaboration with data scientists and clinical operations teams.
Quality, Not Quantity
The size of the available database for matching previous trials to new protocols depends on the indication. Non-small cell lung cancer, for example, may have hundreds of studies as potential inputs into the machine learning algorithm, while a less common medical condition may have only five to seven, says Akbari.
The collective amount of trial information is massive—thousands of studies—but for the purposes of site selection “the volume of the data is not always the determinant,” he continues. “It is the quality of the data that you have and the similarity of selected sample to the actual trial of interest. So, five to seven good studies usually give you the same output as 100 or 110 studies that may be of mixed quality.” Even for lung cancer studies, good site selection can be accomplished with perhaps seven to 10 prototypical trials.
It’s a bit like political polling, offers Akbari. “You don’t sample the entire state... [but] the right sample that you think is representative of the population.”
Editor’s note: Bardia Akbari’s SCOPE presentation is geared toward individuals who are interested in learning about a comprehensive set of mechanisms for protocol evaluation and selection of countries and sites for better performance and enrollment predictions.