Factors predicting access to medications for opioid use disorder for housed and unhoused patients: A machine learning approach

Background Opioid use disorder (OUD) is a growing public health crisis, with opioids involved in an overwhelming majority of drug overdose deaths in the United States in recent years. While medications for opioid use disorder (MOUD) effectively reduce overdose mortality, only a minority of patients are able to access MOUD; additionally, those with unstable housing receive MOUD at even lower rates. Objective Because MOUD access is a multifactorial issue, we leverage machine learning techniques to assess and rank the variables most important in predicting whether any individual receives MOUD. We also seek to explain why persons experiencing homelessness have lower MOUD access and identify potential targets for action. Methods We utilize a gradient boosted decision tree algorithm (specifically, XGBoost) to train our model on SAMHSA’s Treatment Episode Data Set-Admissions, using anonymized demographic and clinical information for over half a million opioid admissions to treatment facilities across the United States. We use Shapley values to quantify and interpret the predictive power and influencing direction of individual features (i.e., variables). Results Our model is effective in predicting access to MOUD with an accuracy of 85.97% and area under the ROC curve of 0.9411. Notably, roughly half of the model’s predictive power emerges from facility type (23.34%) and geographic location (18.71%); other influential factors include referral source (6.74%), history of prior treatment (4.41%), and frequency of opioid use (3.44%). We also find that unhoused patients go to facilities that overall have lower MOUD treatment rates; furthermore, relative to housed (i.e., independent living) patients at these facilities, unhoused patients receive MOUD at even lower rates. However, we hypothesize that if unhoused patients instead went to the facilities that housed patients enter at an equal percent (but still received MOUD at the lower unhoused rates), 89.50% of the disparity in MOUD access would be eliminated. Conclusion This study demonstrates the utility of a model that predicts MOUD access and both ranks the influencing variables and compares their individual positive or negative contribution to access. Furthermore, we examine the lack of MOUD treatment among persons with unstable housing and consider approaches for improving access.