OBJECTIVES: The NarxCare score is an AI/ML-based clinical decision support system intended to measure opioid overdose risk across 46 states’ prescription drug monitoring programs (PDMPs). Concerns persist regarding potential bias in NarxCare and how this impacts equitable opioid prescribing. Our study aimed to scrutinize key predictive features used in NarxCare's prediction, focusing on unveiling likely discriminatory patterns.

METHODS: While NarxCare scores are constructed from multiple features, two features are particularly important: "multiple prescribers" and "overlapping prescriptions.” The former captures patients with three or more unique prescribers throughout the study window, and the second pinpoints overlapping prescriptions for more than seven days within 1-month intervals. Our analysis utilized California's PDMP data spanning from 2010 to 2022 and 5-digit ZIP Code-demographic variables obtained from the US Zip Codes Database (Pareto SoftwareTM, version 2023). We employed the appropriate t-test and systematically examined the associations between patients' neighborhood characteristics and the prevalence of "multiple prescribers" and "overlapping prescriptions.”

RESULTS: Our results indicate that the “multiple prescribers” feature may inadvertently demonstrate discriminatory patterns, disproportionately affecting the access to opioid prescriptions in specific neighborhoods among older adults, lower education levels, those with limited economic means, renters, persons with disabilities, and individuals of white, Pacific, Native, and Hispanic backgrounds. Likewise, “overlapping prescriptions” were notably prevalent in older, less-educated, low-income, and unemployed communities, renters, and disabled individuals. Neighborhoods with higher levels of Asian and Hispanic populations showed a lower likelihood of reporting overlapping prescriptions.

CONCLUSIONS: While the complete list of features and their respective weights in NarxCare's proprietary algorithm remains unpublished, our investigation revealed biased predictions associated with two key features. It emphasized the importance of transparency and the imperative of thoroughly examining NarxCare's proprietary algorithm.