Machine Learning is Gaining Relevance in Predicting Trial Performance, Experts Say
Sites and sponsors are starting to gear up their use of machine learning algorithms to predict key elements in clinical trials, but successful integration relies on finding the right indicators to predict trial performance.
“There are enough indicators out there to do an accurate prediction,” said Elvin Thalund, director of industry strategy at Oracle Health Sciences, but the key is finding the historical data to support those indicators.
The indicators include things such as the disease or condition being studied, study population, eligibility criteria, subject numbers and treatment duration. In addition to indicators identified in the study protocol, Thalund said that historical data from other clinical trials can also offer insight into leading predictive indicators.
Thalund offered an example of a site model that used several leading features/indicators known to impact a study timeline, including the number of countries involved in the study, the number of sites, the experience level of the principal investigators, the trial’s therapeutic area and phase, country code and whether there will be central or local approval of the protocol.
The indicators were fed into a machine learning algorithm, which predicted that it would take 101 days to activate a site. Although this was a proof-of-concept model, Thalund suggested that machine learning may offer real-world value in helping sponsors predict the time it will take between contacting sites and completing site activation.
In addition, Thalund noted that machine learning data may be helpful for gauging enrollment. If a machine learning algorithm identifies certain factors in the trial design that may lead the study to fall short of enrollment goals, such as a treatment protocol that presents a significant burden to participants, sites can then alter the protocol enrollment rate requirements prior to protocol approval, which may likely improve sites hitting their enrollment milestones.
Mary Jo Lamberti, of the Tufts Center for the Study of Drug Development (CSDD), told CenterWatch Weekly that machine learning and natural language processing are frequently applied in clinical research to aid patient recruitment and selection while assisting in clinical trial design. According to CSDD, approximately 42 percent of pharmaceutical and biotechnology organizations reported using machine learning and natural language processing for patient recruitment and selection.
Lamberti mentioned that chatbots were consistently being used by respondents in conjunction with machine learning to identify patient engagement. According to Rob Scott, chief medical officer at AbbVie, the use of chatbots for virtual engagement in decentralized trials can be used to collect information about how patients are doing during the trial. The artificial intelligence (AI) interactions can then be fed into a machine learning algorithm to predict which patients are engaging and which patients are likely to stop compliance or dropout of the trial altogether. Sites can then engage with these patients to find strategies to mitigate potential dropout, such as offering advice on how to manage their participation or improving flexibility in visit schedules.
“If you’re collecting data using electronic patient-reported outcomes, like a mobile app that can track patient behavior, machine learning can also be used to predict dropouts,” said Scott. Data on the number of times patients log into an electronic data-capturing system or mobile app and how many times patients check certain information within these systems can ultimately be used to gauge study engagement and potential dropouts with machine learning techniques.
Jeff Kasher, president of advocacy group Patients Can’t Wait, added that a behavioral analysis as well as socioeconomic factors can be used to identify who will be likely to enter a clinical trial and who will be retained. “These factors can be put into an algorithm and machine learning can be applied to get better insight on how likely patients will enroll on time, where you’re going to potentially lose patients and the profile of patients who will likely adhere to the trials,” he said.
Sponsors and sites looking to incorporate machine learning in planning and conducting their trials should collaborate across the industry to collect enough quality data for their indication, Thalund said, adding that there exists a need for industry standards to support data normalization and anonymization so that the industry can work together to create the needed data to perform these predictions of clinical trial timelines and outcomes.