Kieran Richards
7th May 2025
In an ageing world people are increasingly living with multiple long term health conditions, and consequently are being prescribed multiple medications more often. Taking multiple medications concurrently, also known as polypharmacy, is associated with several negative impacts such as increased risk of adverse side effects, increased risk of drug-drug interactions, and reduced adherence to prescription instructions. Historically, discussion around polypharmacy has centred around the number of drugs a person is taking. However, other authors have emphasized the need to broaden the focus to other aspects of polypharmacy; highlighting that in some cases many concurrent drugs are appropriate and provide benefits that outweigh the risks.
This work proposed a new, probabilistic, clustering model which can identify profiles of prescription patterns based on four key aspects of each medication for individuals in that cluster:
The proposed model leverages Bayesian probabilistic methodology to incorporate uncertainty directly into the model. An individual is not assigned to a single cluster but instead is described by the likelihood they belong to each cluster. In this way, the same drug can be highly likely in multiple clusters which describe how that drug might appear in different polypharmacy patterns. Moreover, the Bayesian approach enables the model to natively handle the common left-censoring in the data. Often the data describe that an individual started taking a medication before a particular date, but the precise start date is unknown. Here we can apply data augmentation, incorporating the limited information that is available whilst acknowledging the lack of precision in the model’s uncertainty.
Applied to a small dataset of 4,600 individuals in CPRD GOLD with Stage 5 Chronic Kidney Disease, the model identified 20 clusters. These 20 clusters distinguished individuals on the basis of the four facets, potentially providing a useful tool for exploring polypharmacy beyond the number of medications. Furthermore, the model presented in this work has laid the foundation for ongoing work that offers improvements including simpler temporal modelling, greater model scalability, and improved interpretability of each aspect’s contribution.