Using Medication Dispensation Data to Identify Clusters with Similar Prescribing Patterns in Older Adults Living with...

Older adults living with dementia are a heterogeneous group, which can make studying optimal medication management challenging. Unsupervised machine learning is a group of computing methods that rely on unlabeled data-that is, where the algorithm itself is discovering patterns without the need for researchers to label the data with a known outcome. These methods may help us to better understand complex prescribing patterns in this population. The objective of our study was to use clustering methods to determine whether common prescribing clusters exist in older adults newly identified as living with dementia in Ontario, Canada and to examine the association between individual clinical and demographic characteristics and those clusters.

Data were derived from population-based health administrative databases, including medication dispensation data. The hierarchical clustering algorithm started with each individual and merged individuals with the most similar prescribing patterns into a group, continuing this process stepwise until only one cluster remained. The optimal number of clusters was selected through clinical review and fit statistics. We examined the association between individual characteristics and prescribing clusters using bivariate multinomial models.

In 99,046 individuals living with new dementia, we identified six prevalent clusters of individuals with common medication subclass patterns: higher dispensation of angiotensin-converting enzyme-specific cardiovascular (22.6% of the population), central nervous system-active (21.3%), hypothyroidism (22.9%), respiratory (3.9%), and angiotensin receptor blocker-specific cardiovascular (6.1%), as well as a group with lower dispensation of medications in general (23.1%). Specific demographic, clinical, and health-service-use characteristics were associated with assigned clusters.

Within individuals living with dementia, prescribing clusters reflected meaningful differences in clinical and demographic characteristics. The results suggest that applying clustering methods to pharmacological data may be useful in estimating complex comorbidity patterns to better describe a heterogeneous population of people living with dementia. Future studies could examine whether these clusters better predict health service use, disease progression, or medication-related adverse events compared with other measures.