Spectral Clustering Identifies High-risk Opioid Tapering Trajectories Associated with Adverse Events

MONIKA RAY, Joshua J. Fenton, Patrick Romano

Keywords: high dose opioids, spectral clustering, patient subpopulations, personalised medicine, healthcare, opioid crisis, phenotype clustering

TL;DR: Spectral Clustering Identifies High-risk Opioid Tapering Trajectories

Abstract: National opioid prescribing guidelines and related quality measures have stimulated changes in opioid prescribing. Studies have shown that rapid dose tapering may be associated with increased opioid-related and mental health events in some patient groups. However, we do not know enough about the trajectories of dose tapering implemented in clinical practice, and how heterogeneous populations of patients respond to different treatments. Our aim was to examine prescribed opioid doses in a large, longitudinal, clinically diverse, national population of opioid-dependent patients with either Medicare or commercial insurance. We performed phenotype clustering to identify unsuspected, novel patterns in the data. In a longitudinal cohort (2008-2018) of 113,618 patients from the OptumLabs Data Warehouse with 12 consecutive months at a high, stable mean opioid dose ($\geq$50 morphine milligram equivalents), we identified 30,932 patients with one dose tapering phase that began at the first 60-day period with $\geq$15\% reduction in average daily dose across overlapping 60-day windows through seven months of follow-up. We applied spectral clustering as we preferred an assumption-free approach with no apriori information being imposed. Spectral clustering identified several cluster-cohorts, with three that included over 98\% of the sample. These three clusters were similar in baseline characteristics, but differed markedly in the magnitude, velocity, duration, and endpoint of tapering. The cluster-cohort characterised by moderately rapid, steady tapering, most often to an end opioid dose of zero, had excess drug-related events, mental health events, and deaths, compared with a cluster characterised by very slow, steady tapering with long-term opioid maintenance. Moderately rapid tapering to discontinuation may be associated with higher risk than slow tapering with longer-term maintenance of opioid analgesia. Furthermore, several clusters highlighted a cohort that had complete taper reversals indicating a treatment failure as the tapering was not maintained. Our findings suggests that identifying subtle yet clinically meaningful patterns in opioid prescribing data, such as patterns within the dose trajectories, can highlight the distinct characteristics separating subpopulations.