(49) Using Interpretable Machine Learning to Predict Access to Buprenorphine in Pharmacies

Background & Introduction: Limited availability of buprenorphine products in community pharmacies is a significant barrier to continuity of care for persons in treatment for opioid use disorder. Recent, multi-state secret shopper audits of community pharmacies have found that 48.3% and 57.9% of pharmacies were able to dispense buprenorphine in response to requests to patients and providers, respectively. Limited availability of buprenorphine raises questions of socioeconomic disparities in availability. The US Census Bureau’s American Community Survey offers a vector of features that may be used to quantify social determinants of health within small population units. Two indices of features from the American Community Survey used to examine the availability and quality of care across socioeconomic strata include the Area Deprivation Index and the Social Deprivation Index. While these indices provide a convenient way to rank the residents within an area along socioeconomic strata, their scores alone do little to characterize the residents within that area. In this analysis, we applied the unique features that comprise both indices to predict availability of buprenorphine in community pharmacies using an interpretable machine learning approach. Our aim was to provide an accurate predictive model with readily understandable rules that determine where buprenorphine may be located within a community.

Methods: The results of Hill and colleagues telephone audit of buprenorphine availability in 4,984 community pharmacies in eleven states were used to define the dependent variable in this analysis, a binary indicator for whether a pharmacy was willing to dispense fourteen buprenorphine/naloxone films in response to a telephonic patient request. Pharmacies were geocoded to the census block group level. Next, 23, unique features of the Area Deprivation Index and Social Deprivation index were extracted from the 2022 American Community Survey at the census block-group level to characterize the demographic, employment, income, and transportation characteristics of the area surrounding the pharmacy. Rural urban commuting area code and pharmacy type (chain or independent) were also included as features. Random forest modeling was used to to predict buprenorphine availability. Model tuning prioritized parsimony and interpretability by limiting depth to eight levels and requiring a minimum node size of 50 observations. Next, a simplified tree ensemble learner (STEL) was used to extract a set of pruned decision rules from the full model. Rules with an importance threshold greater than 5% were ignored. Accumulated local effects (ALE) analysis was used to understand the contributions of individual features within the selected rules toward the final predictions

Results: The random forest predicted buprenorphine availability with a positive predictive value of 94.2%, a negative predictive value 77.0% and an overall accuracy of 89.1% (95% CI: 88.0-90.2%). The three most important predictors of buprenorphine availability, determined by Gini impurity, were the proportion of foreign born residents, the proportion of high needs individuals (children under the age of five, females between the ages of 15 and 44, and adults over the age of 65), and the proportion of Hispanic or Latino/e/a residents residing within a block group. ALE analysis demonstrated that each of these values were associated with reduced probability of buprenorphine availability. STEL analysis was used to rank multi-feature rules that minimized predictive error. STEL demonstrated that chain pharmacies in areas with low to moderate white-collar employment (less than the eighth-decile), and a lower proportion of foreign-born residents (less than the seventh-decile) were more commonly predicted to stock buprenorphine. Conversely, pharmacies in areas with lower median household income (less than the sixth decile), more foreign-born residents (fifth through ninth deciles), and more single parent households (fifth through ninth decile) were less commonly predicted to stock buprenorphine.

Conclusion & Discussion: Buprenorphine availability in community pharmacy can be accurately predicted by the socioeconomic and demographic characteristics of the residents of the census block group containing that pharmacy. Our findings suggest that non-white and immigrant populations are less likely to have access to medication for opioid use disorder within community pharmacies. These results, to some extent, may represent diminished demand for treatment within non-white and immigrant communities. Still, limited access within these areas represents a threat to the health and wellbeing of those attempting to access treatment. Our findings are suggestive of a disparity in access but more work is needed to be able to adequately measure treatment demand at the neighborhood level.

References: 1. Hill, L. G., Loera, L. J., Torrez, S. B., Puzantian, T., Evoy, K. E., Ventricelli, D. J., Eukel, H. N., Peckham, A. M., Chen, C., Ganetsky, V. S., Yeung, M. S., Zagorski, C. M., & Reveles, K. R. (2022). Availability of buprenorphine/naloxone films and naloxone nasal spray in community pharmacies in 11 U.S. states. Drug and alcohol dependence, 237, 109518. https://doi.org/10.1016/j.drugalcdep.2022.109518

2.Weiner, S. G., Qato, D. M., Faust, J. S., & Clear, B. (2023). Pharmacy Availability of Buprenorphine for Opioid Use Disorder Treatment in the US. JAMA network open, 6(5), e2316089. https://doi.org/10.1001/jamanetworkopen.2023.16089

3.Deng, H. (2019). Interpreting tree ensembles with intrees. International Journal of Data Science and Analytics, 7(4), 277-287.

4.Faysal, J. A., Noor-E-Alam, M., Young, G. J., Lo-Ciganic, W.-H., Goodin, A. J., Huang, J. L., Wilson, D. L., Park, T. W., & Hasan, M. M.. (2023). An Explainable Machine Learning Framework for Predicting the Risk of Buprenorphine Treatment Discontinuation for Opioid Use Disorder. https://doi.org/10.1101/2023.11.02.23297982