Development of a Machine Learning Model Predicting Response to Aripiprazole Once-Monthly in Patients Diagnosed With Bipolar I Disorder

This abstract describes the development of a machine learning model to identify baseline factors predicting response to aripiprazole once-monthly (AOM) in patients diagnosed with bipolar I disorder (BP-I). Input data were from patients in a multi-phase trial (NCT01567527; Phase A, conversion to oral aripiprazole; Phase B, oral aripiprazole stabilization; Phase C, AOM stabilization; Phase D, randomization to continue AOM [n=132] or switch to placebo [n=133]). The primary endpoint was recurrence of any mood episode in Phase D (AOM, 35 events; placebo, 68 events). Univariate analysis was used to identify potentially relevant variables, including: demographics; medical/disease history; vital signs; clinician-/patient-reported rating scale scores at Phase D entry and changes from Phase B entry. Of 212 variables considered, 34 and 38 were carried forward for AOM and placebo, respectively. Multivariate predictive modeling was performed using ensemble approaches. The AOM and placebo models demonstrated strong performance (AUC/specificity/sensitivity in out-of-sample validation: AOM, 0.83/0.79/0.82; placebo, 0.85/0.78/0.86).
Potential effect modifiers for AOM response included items/domains for quality of life (QoL), symptoms, functioning, and the lifetime number of prior depressive episodes. Initial findings support the potential utility of disease chronicity for predicting response to AOM. The identification of patient-reported QoL as a potential effect modifier highlights the importance of the subjective patient experience. The model identified potential factors that predict response to AOM, and may be used to support clinician-led monitoring and interventions, where needed (e.g., psychoeducation, therapy, medication adjustment).