(#65) Development of a Machine Learning Model Predicting Response to Aripiprazole Once-Monthly in Patients Diagnosed With Schizophrenia

Abstract: This abstract describes the development of a machine learning model to identify baseline factors predictive of response to aripiprazole once-monthly (AOM) in patients diagnosed with schizophrenia. Input data were from 433 adult patients enrolled in a multicenter, open-label, naturalistic, mirror-image study that compared hospitalization rates in the 6 months before and after switching from oral aripiprazole to AOM (NCT01432444). Univariate analyses screened for potentially relevant variables, encompassing demographic characteristics, vital signs, medical history, and data from clinician-/patient-reported outcome measures. Of 163 variables considered, 37 were carried forward. Predictive modeling was performed using an ensemble approach, with steps to address missing data and the overall low frequency of hospitalization events in the trial (8.8%). The model demonstrated strong performance. AUC-ROC was 0.97 overall and 0.86 in out-of-sample validation (specificity, 0.84; sensitivity, 0.78); corresponding values for AUC-PRC were 0.89 and 0.44 (~6-fold greater than that expected from random chance), respectively. Based on preliminary model outcomes, items on the Positive and Negative Syndrome Scale (PANSS), Quality of Life Scale, and Subjective Well-Being under Neuroleptic Treatment-Short Form questionnaire were identified as potential predictors of response to AOM. Preliminary findings support the potential utility of current symptom severity, measured using the PANSS, for predicting response to AOM. Measures of quality of life/well-being related to social engagement and subjective patient-reported experience also appear relevant. The predictive model can be used to inform in-clinic opportunities to monitor factors that predict response to AOM, with subsequent targeted interventions as required (e.g., psychoeducation, therapy, medication adjustment).

Short Description: A machine learning model to identify baseline factors predictive of response to aripiprazole once-monthly (AOM) in patients diagnosed with schizophrenia was developed using baseline demographic and disease characteristics from 433 adults enrolled in a mirror-image trial. Predictive modeling of selected baseline variables showed strong performance. Preliminary findings identified disease symptomology and quality of life as potential predictors of response to AOM, and highlighted the value of the subjective patient experience.

Name of Sponsoring Organization(s): Otsuka Pharmaceutical Development & Commercialization Inc. (Princeton, NJ, USA) and H. Lundbeck A/S (Valby, Copenhagen, Denmark).