Cluster Analysis to Detect Patterns of Drug Use from Routinely Collected Medical Data

Appropriate drug prescription for an increasingly ageing and multi-morbid population can be a challenge for general practitioners. This study uses unsupervised learning methods to identify different types of patient profiles which could inform policymakers and regulators about patterns of drug use, and identify specific clusters of users with unknown drug effects (risk and benefit). Hard and soft clustering methods are proposed to detect patterns of medication use by patients and to estimate the probability of belonging to a certain patient profile. Results showed the presence of expected as well as a surprising patient profile based on fracture risk factors. Challenges associated with unsupervised learning using electronic medical record data are described and an approach for evaluating models in the presence of unlabeled data using internal and external cluster evaluation methods is presented, such that it can be extended to other unsupervised learning applications within healthcare and beyond. To our knowledge, this is the first study proposing cluster analysis for detecting drug utilisation patterns from electronic healthcare records in the routinely-collected SIDIAP database.