There are many reasons why patients might be nonadherent – whether
it’s study fatigue, forgetfulness, or juggling other responsibilities
and stressors. Despite the best efforts of site teams to identify the
most qualified and motivated patients, ensuring treatment plan
compliance over the course of an entire study has continued to impact
data quality and study performance. Site leaders and sponsors alike have
often pondered what they could do to predict patient behavior without
the aid of a crystal ball.
Today, they no longer need to wish for the aid of a fortune-teller to
help optimize their studies. Knowing that an individual’s future
behavior is likely to be consistent with their past behavior, sponsors
can predict adherence and even the likelihood that a patient will drop
out with high accuracy. Using modern technologies powered by
sophisticated artificial intelligence, we can gain insight into patient
behaviors before a trial begins and optimize cohorts by selecting
participants with a high likelihood of compliance to the medication
regimen.
Can We Tell the Future? Let’s Test It
Using our dataset of dosing behavior of a million doses, we built a
predictive model to do a number of things. The first is a binary
prediction to determine if a patient will be above or below a particular
adherence threshold at the end of the trial. The second is to predict
the patient’s actual adherence at the end of the trial, and the third is
to see if adherence can be used as a good predictive indicator of the
likelihood of dropout.
Because of the unique data set we hold, the models are able to
utilize rich information on patient engagement, not just simply if a
dose was taken. Some of the factors included in the model include how
long it takes a patient to dose, variations in the time of day they
dose, or how long after their alarm time they dose, combined with
information such as study duration and therapy area.
So, does it work? Yes! The models perform with up to 88% accuracy in predicting both a binary adherence threshold and a patient’s actual end adherence using as little as 14 days of data, and excitingly, the accuracy of the models using as little as five days of data is around 80%.
What This Means for Future Studies
The ability to accurately predict a patient’s adherence opens up a number of opportunities to decrease risk in our trials. Let’s explore some of the most impactful.
- Randomization strategies: An understanding of how engaged a patient is likely to be with their treatment regimen provides sponsors with the opportunity to do a number of different things. Eligibility criteria can be specified to enroll only the patients who are most likely to be adherent over the course of the trial, or if sponsors prefer, the information can be used to stratify the patients to ensure a balance across treatment arms.
- Patient specific support: If sites are able to be notified of which patients are likely to struggle the most to remain on track with their treatment, they can provide additional support to those patients and maintain line of sight into their behavior, gaining insight into any potential real-time or future adherence issues. This allows them to create patient-centric interventions that provide personalized support and encouragement, focusing their efforts on the patients most in need.
There are many other applications of this type of predictive capability, and we are excited to be working with our customers to explore the ways in which AI and machine learning can change the way we conduct clinical trials.
To learn more about how Boehringer Ingelheim is applying predictive analytics to improve the effectiveness of clinical trials, please click here to view AiCure’s recent webinar “Leveraging Predictive Dosing to Reduce Clinical Trial Risk,” on-demand.
