Machine Learning in Clinical Trials – Contract Pharma

The breadth of data available is continuing to accelerate. With more insights to dive into, machine learning (ML) capabilities have more potential than ever to enable trial sponsors and study teams to make case-by-case decisions around trial design and execution and better predict outcomes for individual trials. Highlighted further below are several specific examples of its use to enhance outcomes in drug development. However, sponsors should note that ML predictive algorithms can help in a wide range of key development activities, including: risk-based monitoring, electronic Trial Master File document language translations, ongoing medical monitoring, disease detection, biomarker identification, portfolio optimization, and much more.  To get treatments to market faster, while successfully working through regulatory updates and other challenges, sponsors are increasingly relying on innovative approaches and technology solutions to overcome long-standing challenges in drug development as well as new ones brought about by the pandemic. As such, sponsors are also equipping their solutions toolboxes with machine learning capabilities to enhance predictive modeling activities that help improve trial operations at any stage of development. Data-driven trial decisionsWe know the clinical development landscape has increased in complexities in recent years, prior to COVID-19. The global pandemic put long-standing challenges in clinical trials on center stage—making sponsors face an influx of difficult choices as they looked for ways to start trial planning or maintain successful trial continuity now and in the future. Those who have access to ML-based technologies and related expertise are able to build effective algorithms to take a deeper look into the predictors that lie in the expansive level of human data insights available, optimizing various stages of trial operations, including the examples below. Trial outcome predictionsA poorly designed study or a suboptimally-positioned molecule can make or break the success of a trial as well as create additional costs and inefficiencies for sponsors. Leveraging healthcare data sets, ML tools can help assess and optimize compound positioning as well as optimize primary and secondary endpoints in study design. Ultimately, through better study design, sponsors and study teams are setting themselves up for more realistic outcomes and potentially less time spent on protocol development and related amendments.  For sponsors having to consider abandoning or deprioritizing a treatment based on failure to meet Phase 2 or 3 endpoints, these technologies can help responding sub-groups or even completely repurpose the molecule, providing a new approach to portfolio rescue. This opportunity is evidenced by the explosion of companies that are using AI to identify portfolio inefficiencies. Site selectionTo highlight how we can use modern data science to drive better site selection, think about how critical it was at the start of the pandemic for sponsors to determine sites where there was access to enough target patients who met inclusion/exclusion criteria, especially given limited site accessibility and the heightened focus on patient safety given site visits and more.  Decentralized trial solutions, such as telehealth and mobile research nurses for at-home labs and other procedures, greatly helped with remote patient participation. However, for site teams and staff, it was still critical to mitigate risks of increasing cost and time for patient recruitment, especially with limited staff, and potential failure all together. At IQVIA, we aimed to decrease these risks by using ML-based algorithms based on real-world evidence, including claims data and COVID-19 case surveillance data, to map patient populations and identify sites with the highest recruitment potential as well as provide insights into what recruitment strategies may be best for the study and patients at hand. This helped sponsors potentially fine tune their strategy and focus to the right sites, reducing the number of sites involved and related risk of under-enrollment.  Within that, and as importantly, using these models can help locate sites that improve patient diversity in trials in terms of race and ethnicity. IQVIA’s methodology to use machine learning to better recruit diverse patient populations focuses on reviewing trial protocol details such as the condition, trial site features, past performance, claims data, and patient demographics at the trial sites to rank sites accordingly. The ML-based algorithm learns to associate these features as the desirable traits for trial sites to then extract or pinpoint optimal trial sites.  Patient population discoveryIn mining granular healthcare data that is de-identified at the source (e.g., retail pharmacies, insurers, electronic medical records), healthcare technology experts can apply ML-based algorithms to better predict patient segments of interest. By helping to identify patients who are undiagnosed, untreated, or at-risk for the studied disease or condition, the resulting improved outcomes help ensure the right therapies reach the ideal patients.  Through ML models, it is possible to examine patterns in patient electronic health records that may not be detected through more traditional approaches. One specific way technologists do this is to identify and use a sample group or “training cohort” of patients who exhibit the specific characteristics of interest to train the ML model to then, apply the process to the full data set to identify patients predicted to display those same characteristics. For some with rare diseases, it can take years to be diagnosed. In some cases, by the point of diagnosis, there may be significant health deterioration and reduced quality of life. It is a prime example of how machine learning can help understand complex patterns, treatments, and more to make connections towards patient identification earlier.  Humanizing decision intelligence As industry experts continuously work to enhance ML solutions to accommodate the needs of tomorrow’s drug development, it is critical to first address any roadblocks that may hold back further usage. It is safe to say that there is a bit of skepticism among clinical trial professionals, investigators, and others in the industry on bringing in machine learning solutions to help enhance ways of working they are comfortable with and know is effective. It is the job of data scientists and technologists to have realistic expectations of the end user and help alleviate their skepticism by considering human side of our work when developing tech-enabled solutions. Users of ML-based predictive algorithms and related recommendations are people, and it is critical to keep in mind that people do not make decisions like machines. Leaning into the psychology and processes of how decisions are made and creating a user experience can help optimize decisions made by using ML-based predictors. By applying decision intelligence, technologists aim to understand the how’s and why’s of users’ decisions to then augment approaches to ensure information needed to make optimal decisions are there. Unlike other industries, decisions made in clinical research are very critical, potentially life-saving in some cases. It’s very important to help ensure decisions are well-informed for the best outcomes possible.  From creating efficiencies to enhancing patient-centricity to predicting complex trial outcomes and more, effective ML-based algorithms and applications are allowing sponsors to navigate a complex landscape with smarter decisions across the spectrum of trial activities. As key industry stakeholders better understand the varying benefits of these predictive outcomes solutions, the potential for further use in vital drug development will get stronger by the day. That said, it is equally important for those on the forefront of development to help ensure growing acceptance of these solutions by always taking the human factor of decision making – even in those decisions involving AI – into consideration. 

Lucas Glass is the Vice President of the IQVIA Analytics Center of Excellence (ACOE). The ACOE is a team of over 200 data scientists, engineers, and product managers that research, develop, and operationalize machine learning and data science solutions within the R&D space. Lucas has launched over a dozen machine learning offerings within R&D such as site recommender systems, trial matching solutions, enrollment rate algorithms, drug target interactions, drug repurposing, molecular optimization. Lucas’ machine learning research which is dedicated to R&D has been accepted at AAAI, WWW, NIPS, ICML, JAMIA, KDD, and many others.