By Deborah Borfitz 

July 27, 2021 | The potential of artificial intelligence (AI), internet-connected devices, wearables, and cloud computing to disrupt traditional clinical trials was explored during a presentation on patient-centered endpoints at the recent DIA 2021 Global Annual Meeting. The common goal is to make studies more palatable for participants, improving their engagement and retention, and to help pick up the pace and cost of making new medicines, according to Susant Mallick, who leads healthcare and life sciences professional services in the EMEA region for Amazon Web Services (AWS). 

As it is, 30% of people who enroll in studies end up dropping out and 40% stop adhering to protocol requirements after five months, Mallick says. Seventy percent of participants live more than two hours from the study site. They will be more effectively engaged as industry moves toward personalized messaging and virtual clinical trials using IoT devices. 

Engaging patients both at home and at study sites has become “quite common,” says Nagaraja Srivatsan, chief digital officer of technology solutions at IQVIA. That alone has helped trials become more patient- and site-centric.

Engagement efforts need to extend to the post-trial period to inform future study design and technology should be directed at reducing the burden of participation, he says. With laws protecting patient privacy and data strengthening, engagement solutions likewise need to make privacy a priority. 

IQVIA has recently worked on several COVID-19 vaccine trials, a few of which took less than 60 days to move from protocol to first patient in and reached database lock in six months, Srivatsan shares. AI and targeted technology models were used up front to identify the right patients and activate the right sites. 

Engagement methodologies included a process to identify appropriate patients via trial matching, profiling, and referral, as well as a direct-to-home model that included the use of telemedicine, electronic clinical outcome assessment tools, and in-home nursing or other human touchpoint to help ensure protocol compliance, he says. Notably, IQVIA is doing “data tokenization” post-trial to look at data in the real world and understand how different drugs and therapeutic solutions are working. 

“Technology has to be intuitive for patients and sites, it has to be intelligent—which is enables by data and machine learning and AI—and, more importantly, it has to be interoperable,” says Srivatsan. 

Meeting Needs 

Machine learning can be used in any clinical trial to manage patient engagement more efficiently and effectively, says Jacob Sunol, director of the Danish consultancy Lighthousing ApS, who discussed its use in oncology. The first step is to design a trial for success, which typically involves using a dataset to simulate the study and test the inclusion criteria. Claims data, for example, might be used to find patients who have switched from a first- to second-line treatment and to learn how long it takes for the switch to happen. 

A machine learning technique called clustering can be used to create groups and, if there are too few patients with the desired features, it will be necessary to define how the group can be effectively amplified, he says. One option is federated learning on hospital data—meaning, “training models into each hospital’s data and then pulling the model out” so the data never leaves the institution’s protective custody. The approach facilitates dialogue with hospitals’ chief information officers and allows data from the different hospitals to run separately in the cloud and be aggregated together later. 

Once a clinical trial is running, investigators need to be provided with the right tools for the job, Sunol says. Many machine learning models can help annotate tumors, for example, and cloud providers offer pre-defined natural language processing models that can be used to read the notes of healthcare providers or transcribe their recorded speech into text. 

Another possibility is to build a digital biomarker leveraging data from previous clinical trials to, for example, understand the progression of a tumor and use that to create predictive models and adapt the study to patient needs, says Sunol. 

User personas could be developed to represent patients at different points in their therapeutic journey (e.g., newly diagnosed, patient experts), he continues. The user experience can also be personalized with machine learning, and not just using traditional A/B testing where patients get randomized into one platform or the other. Personalization can happen even among patients in the same disease area and on the same treatment.

To deal with adherence problems in studies, it may not be enough to send out reminders or have nurses call participants because they may “not want to remember… and will cheat,” says Sunol. When this happened in a study he was working on, “we changed the strategy and made them learn about the benefit of injecting their drug.” This sort of knowledge can be gained during data aggregation and clustering, he adds. 

“In many organizations there is a conflict between biostatisticians and data scientists that we need to tackle,” Sunol points out. The problem emerges from the fact that data collected in a clinical trial has to be analyzed and side effects reported, which means “extra problems for biostatistics team. I always recommend that both sit together because one without data will not be able to run the models and one with a lot of data will have to do 10 times more work.” 

Even after a study ends, post-trial responsibilities continue, he concludes. The many learnings from a clinical trial can be applied later to digital solutions for patients, as well as informing the development of future protocols. 

Privacy By Design 

In research and development, the use cases for RWD include synthetic control arms, new indication finding, patient journey mapping, comparative effectiveness, and benefit-risk assessment, according to Harini Gopalakrishnan, head of technology strategy at Sanofi. Creation of a synthetic control arm is the “most impactful” because it shortens clinical trial duration. 

When AI and AI-based techniques are being leveraged to mine the data, as is increasingly the case, the result is often a trial with a better validated outcome, she adds. AI can also be deployed for patient data management, including privacy preservation. “The key differentiator for companies is the extent to which they can generate insights and evidence from multiple data sources.” 

Using AI at scale is still challenging, Gopalakrishnan says, because of the necessity of hiring many data scientists and the change management required. When claims and electronic health data are being licensed from a company like IQVIA or Flatiron Health, the data being purchased “is probably already of good quality.” 

But with internally-generated data, an assortment of data management, quality, and stewardship issues emerge, she continues, and one of the most important when using AI is ensuring patient protection via “privacy by design” principles. Technology can help enforce the “guardrails,” so patients do not get re-identified without their specific consent. 

Sanofi is currently piloting an AI algorithm for “differential privacy” that masks key identifiable fields on demand when, for example, the company wants to analyze legacy clinical trial data for a secondary purpose, Gopalakrishnan shares. Privacy by design is important for pharmaceutical companies both in maintaining a trusted relationship with patients and fulfilling their obligation to ensure extracted insights are backed by the proper data propagation and management methodologies and are thereby reproducible.

“Any key system that you architect today if you are dealing with patient data should have at least four pillars,” she concludes. These include “great data management… at speed,” the ability to run various kinds of analytics with point-and-click interfaces, ensuring access control, and an approach to traceability and auditing that enables any analysis to be reproduced—even if it is AI-based. 

“The cloud is the common denominator for all the transformation happening in the industry,” says Mallick, adding that “privacy by design” is not a choice but a mandate. AWS, for its part, is helping customers ensure data storage and usage aligns with prevailing regulatory frameworks, including the Health Insurance Portability and Accountability Act in the U.S. 

Customers can even choose the geographic regions where their data get stored and used, Mallick says. AWS will neither move nor replicate content outside of those chosen regions without a customer’s consent.