USING ARTIFICIAL INTELLIGENCE/MACHINE LEARNING AND REAL-WORLD DATA TO IDENTIFY PATIENTS WITH EARLY STAGE BIPOLAR-I DISORDER FOR WHOM AN ARIPIPRAZOLE MONOHYDRATE LAI IS LIKELY TO BE A FAVORABLE TREATMENT OPTION

Background

Bipolar-I disorder (BP-I) is a chronic psychiatric condition marked by significant functional impairment and clinical management challenges, typically requiring sustained maintenance therapy. 1 While evidence exists to support the benefit of early initiation of long-acting injectable antipsychotics (LAIs) compared to oral formulations, realworld data are lacking to guide personalized treatment decisions. 2 This study aimed to leverage large-scale United States (US) electronic health records (EHR) and claims data and apply machine learning (ML) methodologies to identify demographic, clinical, and individual treatment and timing factors associated with most favorable response to aripiprazole monohydrate LAIs (i.e., aripiprazole once monthly [AOM] and aripiprazole 2-month Ready to Use [Ari 2MRTU]) in early-stage BP-I.

Methods

This prognostic study used real-world data from a large, integrated EHR and open claims database in the Atropos Evidence Network™ covering all sites of care and pharmacy fills for patients in the US from 2022-2025.3 The primary ML approach used targeted minimum loss–based super learning ensembles to incorporate drivers of observed treatment selection and likely outcomes in predicting characteristics of BP-I patients for whom aripiprazole monohydrate LAIs are likely to be favorable treatment option. The model was trained on 70% of the real world-dataset and validated on the remaining 30%. To assess the impact of time to antipsychotic (AP) treatment initiation, the index treatment for each patient was defined as the last AP treatment started within 3 years of initial BP-I diagnosis with ≥1 year of follow-up. The primary treatment success measure was absence of both all-cause inpatient admissions and an AP treatment regimen change (switch or stop) during 12-months post-index period. To test robustness, sensitivity analyses were performed (different datasets, study operational definitions, statistical and ML techniques).

Results

Among the 98,972 (validation sample n=29,510) patients starting AP treatment within 3 years of initial BP-I diagnosis and with ≥1 year of follow-up, mean age was 41 ± 15 years, 62% were female, 33% had recent substance use documented, and the mean time from diagnosis to index treatment was 1.06 ± 0.97 years. The most common index treatment was an oral atypical AP (51%). Overall, 66% of patients stayed free of hospitalization and AP regimen change within 1 year of initiating index AP treatment. The most important outcome predictors were recent patterns of inpatient and emergency department utilization.The final model was able to distinguish between different treatments for BP-I, including between AOM and Ari 2MRTU. The ML policy proposed aripiprazole monohydrate LAIs should be used earlier after diagnosis. Ari 2MRTU demonstrated higher expected benefit over shorter-term LAIs (AOM), but there was still a benefit for both LAIs in this population. Ari 2MRTU had the greatest estimated treatment effect and was preferred for patients within 5 months of initial BP-I diagnosis. The estimated treatment success (if patients had been on the ML policy-proposed optimal treatment) was 74% (95%, CI: 73%-76%). This was significantly better than the overall observed success rate (66%, p < 0.001).

Conclusion

Patients with BP-I are likely to benefit from earlier initiation of Ari 2MRTU. This ML predictive approach should be refreshed in the future as more AP treatment data becomes available.