Predicting Antidepressant Deprescription with Machine Learning Using Administrative Data

The high prevalence of failed antidepressant deprescription attempts makes it difficult for clinicians to identify suitable candidates for discontinuation. In this study, we use the Pharmaceutical Benefits Scheme (PBS) dataset, which contains rich longitudinal dispensing data from Australian primary care and developed supervised models to predict successful deprescription. To overcome the challenge of observational administrative data, we developed two distinct annotation pathways (retrospective and prospective) to identify successful and unsuccessful deprescription cases and trained several supervised machine learning models using the labelled data from each pathway. The best accuracy in predicting successful deprescription was reached using XGBoost (0.81) and Random Forest (0.90) for retrospective and prospective annotation, respectively. The study results demonstrate the potential of administrative healthcare data for developing clinical decision support tools, though further validation against clinical outcomes is needed.