AbstractIntroductionMedication nonadherence is a public health concern and can impact clinical trial data quality. Traditional compliance collection (pill counts, diaries) can be unreliable in central nervous system trials. As such, strategies such as adherence technologies may play a key role in trial outcomes. AiCure, a computer vision-assisted dosing mobile application (app), collects dosing data and connects patients to sites for dosing support. Phone-based computer vision algorithms confirm dosing and transfer videos for artificial intelligence and human review. Boehringer Ingelheim is partnering with AiCure on pilot trials using AiCure adherence data to improve patient retention and clinical trial data quality. Here we report initial findings.MethodsThis pilot used data from two Phase II trials on the efficacy and safety of BI 409306 in people with schizophrenia (NCT03351244) or Attenuated Psychosis Syndrome (NCT03230097). The AiCure mobile app alerted participants to dosing protocols. The dose event was visually confirmed, providing sites a real-time view of adherence and allowing for targeted outreach and intervention. Adherence data from the first 2 weeks generated quantitative, machine-learning models to predict the individual adherence over the trial. Predictive modeling explored different monitoring periods (7-, 10-, and 14-day) and adherence cutoff points (0.8, 0.7, 0.6).ResultsInitial AiCure assessment identified 43% of participants in NCT03351244 as ≤80% compliant (definition of compliance >80% compliant). Variance in adherence rates between electronic case report forms (eCRF; 78%) and AiCure (26%) data was also observed in the highly compliant/adherent group in NCT03230097. Using the first 2 weeks of adherence data (both studies combined), a participant’s adherence predicted their average adherence for the remainder of the trial. Observation of a participant’s adherence for the latest 4 weeks predicted the probability of premature dropout from the trial. There were further correlations of lower predicted adherence with actual disposition-based dropouts.The early adherence predictive model (0.6 adherence cutoff) identified 22%, 20%, and 19% of patients for trial NCT03351244 (total n=235) as high-risk patients (low-adherence prediction) across 7-, 10-, and 14-day monitoring periods, respectively. Of those high-risk patients, 81%, 90%, and 96%, respectively, were truly nonadherent based on actual adherence data. The 14-day monitoring period model provided the lowest false omission rate, indicative of a better performing model.ConclusionsAiCure data provided insights into patient behavior and adherence patterns which would not be available via CRF. Predictive models developed with AiCure adherence data can identify and predict future poor adherers. This creates opportunities to plan interventions and mitigation strategies to improve patient adherence during trials, thereby providing test drugs the best opportunity at proving efficacy.FundingBoehringer Ingelheim International GmbH (NCT03351244/1289-0049 and NCT03230097/1289-0032)
Read full abstract