Unsupervised machine learning identifies opioid taper reversal patterns in a longitudinal cohort (2008–2018)

Chronic pain is commonly treated with long-term opioid therapy, but rapid opioid dose tapering has been associated with increased adverse events. Little is known about heterogeneity in the population of patients on high dose opioids and their response to different treatments. Our aim was to examine opioid dose management and other patient characteristics in a longitudinal, clinically diverse, national population of opioid dependent patients. We used spectral clustering, an unsupervised artificial intelligence (AI) approach, to identify patients in a national claims data warehouse who were on an opioid dose tapering regimen from 2008-2018. Due to the size and heterogeneity of our cohort, we did not impose any restrictions on the kind or number of clusters to be identified in the data. Of 113,618 patients with 12 consecutive months at a stable mean opioid dose of ≥ 50 morphine milligram equivalents, 30,932 had one tapering period that began at the first 60-day period with ≥ 15% reduction in average daily dose across overlapping 60-day windows through 7 months of follow-up. We identified 10 clusters that were similar in baseline characteristics but differed markedly in the magnitude, velocity, duration, and endpoint of tapering. A cluster comprising 42% of the sample, characterised by moderately rapid, steady tapering, often (73%) to a final dose of zero, had excess drug-related events, mental health events, and deaths, compared with a cluster comprising 55% of the sample, characterised by slow, steady tapering. Four clusters demonstrated tapers of various velocities followed by complete or nearly complete reversal, with combined drug-related event rates close to that of the slowest tapering cluster. Unsupervised AI methods, such as spectral clustering, are powerful to identify clinically meaningful patterns in opioid prescribing data and to highlight salient subpopulation characteristics for designing safe tapering protocols. They are especially useful for identifying rare events in large data. Our findings highlight the importance of considering tapering velocity along with duration and final dose and should stimulate research to understand the causes and consequences of taper reversals in the context of patient-centered care.Author summary: Large data warehouses, such as those developed by insurance agencies, allow for population health research which is generalisable. However, due to the enormity and heterogeneity of the data, many hidden or unsuspected patterns are overlooked. These underlying patterns are important because they not only aid in a deeper understanding of the inferences from hypothesis testing but also uncover new testable hypotheses. In our analyses, we show an application of unsupervised machine learning to identify hidden patterns in large data. Our aim was to use methods that did not impose any restrictions on the kind of patterns to be identified nor bias the approach to lead us to results that are well studied in this domain, i.e., we were searching for novel, unreported characteristics. Since large datasets also contain noise, we applied a true data “mining” approach where one can choose to take what is valuable and throw away the rest. Our results using spectral clustering identified clinically meaningful patterns, including rare events such as taper reversals. Our application of clustering to opioid dose tapering is an innovative approach of highlighting new hypotheses to study and also an example of how unsupervised machine learning methods can be used to discover rare events as they do not follow normal distributions. Powerful unsupervised pattern recognition methods can be applied to other population health studies performed on large data to advance scientific knowledge.