Measurement Bias in Machine Learning-Enhanced Opioid Risk Scoring Systems

Over the past two decades, Prescription Drug Monitoring Programs (PDMPs), augmented by machine learning-empowered clinical decision support (CDS) tools, have become essential in monitoring opioid prescriptions. The CDS tool employs proprietary algorithms to evaluate the risk associated with opioid prescribing. However, concerns have emerged regarding potential algorithmic biases that may disproportionately impact marginalized groups. Our analysis of California's PDMP data and ZIP code-based socio-demographic characteristics examined correlations between the common predictive features used in the CDS tools and sensitive attributes. Our findings indicated that predictive features might not favor females, White, Black, Native American, and Pacific Islander populations, older patients, neighborhoods with high percentages of disability and unemployment, and Medicare patients. Given the substantial influence of these CDS tools on prescribing practices, it is imperative to address these biases to ensure equitable access to pain management and to mitigate potential adverse effects on vulnerable patient populations.