Panel Paper: Using Unsupervised Machine Learning to Identify Potentially Problematic Opioid Use in Medicare

Saturday, November 4, 2017
Hong Kong (Hyatt Regency Chicago)

*Names in bold indicate Presenter

Carroline Lobo, Hawre Jalal, Chung-Chou Chang, Gerald Cochran and Julie Donohue, University of Pittsburgh

Background: Potentially problematic opioid use in Medicare has been on the rise. The 2016 Comprehensive Addiction and Recovery Act aims to limit problematic use by restricting the number of opioid prescribers/pharmacies for some beneficiaries via restriction or “lock-in” programs. What constitutes “problematic use” of prescription opioids among the elderly or disabled Medicare beneficiaries is subject to debate. Further, the characteristics of those with potentially problematic use remain unknown. We identify sub-groups of Medicare beneficiaries with potentially problematic opioid use patterns using novel machine-learning techniques which handle complex interactions among variables of interest.

Methods: We identified sub-groups of Medicare beneficiaries based solely on variables constructed using pharmacy claims due to regulatory restrictions on substance abuse diagnosis codes in Medicare data. Our study sample included 190,148 PA beneficiaries enrolled in fee-for-service Medicare from 2007 to 2012 with a new episode of opioid use with continuous enrolment six months prior to the index prescription with no metastatic cancer or no hospice/long-term care use. We used a technique of unsupervised machine learning called k-means clustering analysis. The number of unique opioid prescribers, number of unique pharmacies, and morphine milligrams equivalents per day (MME/day), each measured over 6-month periods – and the mean, maximum, and range of these values across all episodes per beneficiary were used as clustering variables. We compared sub-groups based on their demographic and enrollment characteristics, comorbidities and other medication use and conducted multivariable logistic regression analyses to compare likelihood of all-cause mortality.

Results: We identified five sub-groups based on opioid use patterns. Beneficiaries in sub-group 5 (n=1,467, 0.8%) obtained opioid prescriptions from a maximum of 8.9± 4.1 prescribers and 5.8± 2.3 pharmacies, and had a mean MME/day of 125.4± 123.7. This sub-group was younger (42.9±10.8 years) and had the highest proportion of disabled beneficiaries (96.5%). This sub-group also had the highest prevalence of back pain (95.0%), neck pain (73.1%), and mood disorders (83.3%) compared to other groups. Concurrent use of anti-depressants, anti-anxiety, sedatives/hypnotics, and muscle-relaxants was the highest in sub-group 5. Beneficiaries in sub-group 4 (n=5,223, 2.8%) had a maximum MME/day of 181.6± 173.2. However, the maximum opioids prescribers and pharmacies for this sub-group were much lower than sub-group 5 at 2.1± 0.8 and 1.4± 0.6 respectively. Sub-group 1 accounting for 71.0% of beneficiaries had the lowest utilization among all variables of interest receiving opioid prescriptions from a maximum of 1.3± 0.5 prescribers and 1.0 ± 0.2 pharmacies with a maximum MME/day of 43.8± 39.6. Sub-groups 4 and 5 had a 3.34 [95% CI=2.86, 3.67] and 6.15 [95% CI=5.28, 7.16] odds of all-cause mortality, respectively, relative to sub-group 1.

Conclusion: We identified a small proportion of beneficiaries (<1%), nearly all of whom were disabled, with high numbers of prescribers and pharmacies using high dosages of opioids across episodes of treatment. These beneficiaries also had higher co-morbid conditions and other concurrent health care use. Heterogeneity in opioid use patterns in the Medicare population has implications for efforts to restrict opioid use and improve the quality of care for patients with complex needs.