AI in clinical trial adherence and retention:
a Pandora’s box?

C

linical trial participants adhering to the study protocol and the risk of patients dropping out are key concerns among study sponsors. Current ways to avoid these scenarios often involve labour-intensive tasks. If efforts to avoid these concerns are unsuccessful, they can lead to study failure.

With such concerns, the potential of artificial intelligence (AI) intrigues the clinical trials industry. However, there are still reservations, such as the fear that AI could be a Pandora’s box, unleashing unnecessary risks to clinical trial completion.

One key concern is how AI fits into the regulatory framework. At present, AI is classified as a medical device by the FDA. However, AI is unique from other kinds of medical devices as AI algorithms have an expiration date, in that they cannot be expected to perform the same way in the next decade, explains Massachusetts Institute of Technology principal research scientist Dr Leo Anthony Celi.

“AI is a living medical product that needs to be constantly tweaked and recalibrated,” Celi adds. AI should be separated from medical devices, similar as to how the FDA separates drugs and vaccines, agrees Duke Clinical Research Institute chief science and digital officer Eric Perakslis, PhD.

While this predicament still needs to be addressed, as do many others, sponsors should nevertheless not shy away from their curiosity about using AI to potentially boost patient adherence and retention. Clinical Trials Arena spoke with AI experts who mentioned potential avenues for AI in clinical trials, but also noted blind spots that sponsors should be careful not to miss. Previously, a different set of experts discussed the untapped value of AI in terms of finding the right patients and recruiting them.

During a clinical trial, participants must adhere to a protocol that indicates the dosage and timing of administering a certain medication and must also keep a record of receiving the medication for a prolonged period. However, in some medical conditions such as psychiatric and mental health disorders, there is an inherent risk of adherence failure due to a lack of motivation, notes Digital Health CRC chief innovation officer Stefan Harrer, PhD. Digital Health CRC is an organisation focusing on health innovation and commercialisation that is based in Australia.

AI-driven technology, especially video monitoring systems, can be designed to pick up certain patient behavioural patterns, such as avoidance of taking the medication, and to predict possible non-compliance, Harrer explains. Once certain drug avoidance behaviours are identified, trial coaching technologies can be used to prevent them.

A digital personal assistant that knows about the patient’s trouble with adherence to certain parts of the protocol would alert the participant directly about the challenges they face and provide help or information on how to overcome those obstacles, he adds.

When designing any type of health-related patient monitoring solution, the patient monitoring technology and data regulatory governance strategy need a clearly defined purpose, Harrier adds. Video data is the most sensitive type of data and must be handled with utmost privacy. There is a high bar in video privacy as trust needs to be earned from the user, he explains.

Overall, clinical trials have a very poor record of patient retention, says Harvard Medical School Biomedical Informatics professor and chair Dr Isaac Kohane. Patients can lose interest in the trial, especially when the patient starts to get better.

AI can be used as a customer service management tool and help create a “cult of the study” where the patient feels as they are a part of the wider clinical trial enterprise that helps other people, rather than as if they are a subject needed only for regular blood tests and who is then forgotten after the study ends, Kohane adds.

AI can provide continued updates or newsletters to patients regarding the clinical trial, which is otherwise a labour-intensive part of studies, Kohane says. “It is the feedback to the patients where AI can be extremely effective, because it allows CROs or pharma companies to manage large populations in a cost-effective way,” he adds.

Clinical trial sponsors often use surveys to track patient experience in a clinical trial. However, data is one-dimensional and can miss relevant trends. AI can turn the static datasets into responsive models to track patient feedback, which could help to indicate if the participant is struggling with retention, he explains.

AI is not frequently thought of as an engagement tool, but it has potential to prevent trial attrition, Perakslis adds. For example, if a trial sponsor wants to conduct a monthly survey for a 10-month clinical study, an AI-designed survey would change the set of questions based on the previous answers provided by the participant, he explains.

Still, AI has the potential risk of becoming a “phone tree” if the development of the technology is done in a heavy-handled manner, Kohane says. Over-exposure to constant useless information can cause the oppositive effect of intended AI use and thus disinterest participants.

Constant monitoring of patients with a highly individualised disease is challenging and this high level of oversight is not feasible. Clinicians may not be able to automatically recognise disease issues and immediately adapt to the changes, Harrer says. AI techniques such as machine learning and deep learning can help with patient monitoring and recognise disease patterns in unstructured data collected by wearable devices.

For example, patients with epilepsy usually track the seizure occurrence in a diary, but this approach is unreliable and prone to error in the case of absence seizures, where a patient has no recollection of the seizure. AI techniques can be used to examine collected data and build monitoring technology that can send an alert when a patient has an epileptic seizure, which can be useful in a clinical trial that is investigating a drug that intends to suppress or prevent certain types of seizures, Harrer explains.

AI-assisted monitoring systems help make endpoint detection more efficient, which would result in a shorter clinical trial duration and financial savings, Harrer adds.

Clinical trial data mining via AI

Machine learning can also help clinical trial sponsors to identify cohorts in an existing trial that are worth investigating further, Kohane says. He explains that if a Phase III study failed to meet its endpoints, AI can be used to sift through large sets of data to identify subgroups of patients that benefited from the intervention. This would indicate the opportunity for further engagement with that specific cohort.

AI can play a role in platform trials, Celi says. In such trials, AI can analyse data while the trial is still running and detect potential improved outcomes in certain patients. This can result in a change of randomisation and increased likelihood of a patient receiving a fitting treatment.

Celi notes that there is potential danger if companies decide to protect their AI algorithms under patenting or intellectual property. The inability to examine the AI’s internal workings will make it challenging to interrogate those algorithms.

Perakslis recommends incorporating independent assessments when reviewing AI so that the people who are validating the algorithms are not the same people who wrote them, similarly to how research papers get peer-reviewed. However, he highlights the fact that these technologies will remain somewhat vulnerable to the humans creating them as both are fallible.

With such AI advantages, questions remain about who is going to lead the change. Kohane suggests that CROs are the most obvious player, as they have the largest knowledge base regarding clinical trial cohorts. Once pharma companies understand the scientific and financial value of maintaining patient engagement via AI, they are likely to join. However, in the meantime, AI’s full potential, and even its risks, remain to be seen.