Text

Background/Aim: Polypharmacy is commonly characterised by the number of medications, leaving clinically relevant information such as distinct medication patterns insufficiently understood. Therefore, the objective of this analysis was to identify medication patterns using an unsupervised clustering approach.

Methods: Latent Dirichlet Allocation (LDA) was applied to baseline assessment data from the German National Cohort (NAKO). In the NAKO all medications reported to be taken regularly during the preceding 7 days are included. For the main analysis, corresponding ATC codes were aggregated at level 4. The analysis was restricted to ATC codes reported by ≥ 1.0% of participants and participants taking ≥ 2 medications regularly with complete information on ATC codes and frequency of use. Tuning metrics, model stability, and clinical interpretability were used to determine the number of the clustering topics. For visualisation at topic level, Hellinger-based Multidimensional Scaling plots were used. At participant level, Uniform Manifold Approximation and Projection (UMAP) projection was applied.

Results: 55,052 participants (mean age 57 years, 55% female) fulfilled the eligibility criteria. Highest clinically interpretability was achieved by a 6-topic model which comprised a gynaecological-endocrine hormone pattern, a cardiovascular secondary prevention pattern, a diabetic-cardiorenal pattern, a respiratory/allergy pattern, a male-dominated cardiovascular-urological pattern, and a pattern of chronic pain with psychotropic and neurological co-medication. At the topic level, clinically related cardiometabolic patterns showed a sufficient separation despite their close relationship. At the participant level, compact clusters with dominant topic membership and overlapping regions were visible, aligning with the mixed-membership structure of LDA.

Discussion/Conclusion: In this cohort study, the LDA approach appears appropriate for identifying clinically plausible medication patterns. Based on the main analysis, a 6-topic model is indicated. Sensitivity analyses, profiling of topic-dominant participants, and regression analyses of associated features will follow.