AI-based medication adherence prediction in patients with schizophrenia and attenuated psychotic disorders

Objective: The capacity of machine-learning algorithms to predict medication adherence was assessed using data from AiCure, a computer vision-assisted smartphone application, which records the medication ingestion event.

Methods: Patients treated with BI 409306 were recruited from two Phase II randomized, placebo-controlled trials in schizophrenia (NCT03351244) and attenuated psychotic disorders (NCT03230097). A machine-learning model was optimized to predict overall trial adherence using AiCure data collected over three monitoring periods (7/10/14 days), adherence cut-offs (0.6/0.7/0.8) and timepoints (Start/Mid/End). Area under the curve (AUC), false negative rate, and false omission rate averaged across 10 model cross-validations were analyzed. In NCT03351244, post hoc analyses compared time to first relapse in patients observed as adherent versus those predicted adherent by the model.

Results: Of 235 patients, 60.4 % demonstrated ≥80 % adherence. At an adherence cut-off of 0.8, the 14-day model performed best (AUC: 0.81 versus 0.79 [10-day], 0.77 [7-day]). Within the 14-day model, 0.6 cut-off was optimal (AUC: 0.87 versus 0.85 [0.7 cut-off], 0.81 [0.8 cut-off]). The Trial-End timepoint yielded the most accurate prediction (AUC: 0.92 versus 0.87 [Start], 0.85 [Mid]). Despite NCT03351244 not meeting the primary endpoint, a reduction in risk of first relapse with BI 409306 versus placebo was observed when analyzed with adherent completers (≥80 % across trial; HR = 0.485) and patients with predicted adherence ≥60 % (HR = 0.510).

Conclusions: Adherence data with longer monitoring durations (14 days), lower adherence cut-offs (0.6), and later timepoints (Trial-End) produced most accurate adherence predictions. Accurate adherence prediction provides insights about medication adherence patterns that may help clinicians improve individual adherence.