Algorithmic Fairness in Predicting Opioid Use Disorder using Machine Learning

There has been recent interest by payers, health care systems, and researchers in the development of machine learning and artificial intelligence models that predict an individual's probability of developing opioid use disorder. The scores generated by these algorithms can be used by physicians to tailor the prescribing of opioids for the treatment of pain, reducing or foregoing prescribing to individuals deemed to be at high risk, or increasing prescribing for patients deemed to be at low risk. This paper constructs a machine learning algorithm to predict opioid use disorder risk using commercially available claims data similar to those utilized in the development of proprietary opioid use disorder prediction algorithms. We study risk scores generated by the machine learning model in a setting with quasi-experimental variation in the likelihood that doctors prescribe opioids, generated by changes in the legal structure for monitoring physician prescribing. We find that machine-predicted risk scores do not appear to correlate at all with the individual-specific heterogeneous treatment effect of receiving opioids. The paper identifies a new source of algorithmic unfairness in machine learning applications for health care and precision medicine, arising from the researcher's choice of objective function. While precision medicine should guide physician treatment decisions based on the heterogeneous causal impact of a course of treatment for an individual, allocating treatments to individuals receiving the most benefit and recommending caution for those most likely to experience harmful side effects, ML models in health care are often trained on proxies like individual baseline risk, and are not necessarily informative in deciding who would most benefit, or be harmed, by a course of treatment.