Our analysis utilized California's PDMP data spanning from 2010 to 2022, coupled with 5-digit ZIP Code-demographic variables obtained from the US Zip Codes Database (Pareto Software™, version 2023). Due to health data privacy regulations, the PDMP dataset contains limited patient-level attributes, encompassing age, gender, and payment type. Racial and socioeconomic factors are absent from the PDMP dataset. However, these attributes can be approximated using ZIP-Code-based neighborhood data, encompassing metrics such as educational attainment, median household income, poverty rate, unemployment rate, home-ownership rate, disability status, and racial demographics, including Black, White, Asian, Pacific Islander, Native American, and Hispanic populations. Leveraging both patient-level and ZIP-Code-based characteristics, our study aimed to investigate the correlation between predictive features “multiple prescribers (more than two prescribers),” “multiple pharmacies (more than two pharmacies),” and “daily dosage exceeding 120 MME” features and these sensitive attributes. For each, we divided the data into two groups: patients who exhibited the feature (e.g., received prescriptions from multiple pharmacies or had a daily dosage exceeding 120 MME) and those who did not. ORS documentation1 specifies that these predictive features are evaluated across four-time frames: the most recent two months, six months, one year, and two years. We then calculated the mean values of various “sensitive attributes” within each group for each time frame. These sensitive attributes included age, gender, payment type (Medicaid vs. non-Medicaid), and neighborhood characteristics based on ZIP Codes. We computed a ratio for each feature and time frame by dividing the mean values of sensitive attributes among patients with the feature (e.g., multiple prescribers/pharmacies or high daily dosage) by the mean values among patients without the feature. This approach provided insight into how each feature correlates with sensitive attributes across different time frames. A ratio greater than one signifies a positive correlation between the predictive feature and the sensitive attribute. This indicates that patients characterized by the sensitive attribute are more likely to be flagged by the corresponding predictive feature, as illustrated in Figure 1.
In Figure 1, the predictive features "multiple prescribers" and "multiple pharmacies" show positive correlations with Medicare recipients, older adults, those with a high school education, White communities, the unemployed, and homeowners. Conversely, these features are negatively correlated with Medicaid recipients, individuals with a college education or above, and Asians. The “high dosage” feature is more commonly found among older adults, Native Hawaiians and Pacific Islanders, neighborhoods with predominantly White and Black populations, and homeowners, but is less prevalent among Asians, individuals living in poverty, Medicare recipients, and Hispanic neighborhoods (Figure 1). These findings underscore that the features of multiple prescribers, multiple pharmacies, and high dosages are not neutral indicators of risk. Instead, they are significantly influenced by sensitive attributes, often reflecting systemic inequities. For instance, “multiple prescribers” and “multiple pharmacies” are more frequently observed among older adults, individuals on Medicare, and certain racial and ethnic groups. This pattern may stem from the fact that fewer physicians accept new Medicaid patients compared to Medicare patients 15, and the flexibility of Medicare to request coverage determinations for medications may lead patients to consult multiple prescribers to find someone who better meets their pain management expectations 16. However, the algorithm may misinterpret these patterns as indicative of higher misuse risk, leading to disproportionate penalties for these groups. Moreover, the algorithm’s reliance on these features may result in the flagging of individuals living in predominantly White and Black neighborhoods as higher risk. In comparison, those in Asian and Hispanic neighborhoods may be flagged as lower risk. This can lead to potential undertreatment or unjust scrutiny of the flagged individuals.
The pattern corroborated suspicions raised in multiple opinion papers 4,6,7 regarding ORS’s potential biases. These concerns revolve around how the proxies employed by PDMPs may disproportionately affect marginalized groups, including racial minorities and socioeconomically disadvantaged patients. Our findings revealed that features derived from the purportedly “objective measure” of dispensing records from the PDMP display unfavorable biases towards some sensitive attributes, posing challenges to the predictive algorithm reliant on such biased features.
The United States Food and Drug Administration (FDA) has compiled a list of AI/ML-enabled medical devices that have successfully met the FDA's premarket requirements for overall safety and effectiveness, including evaluating study diversity based on the device’s intended use and technological characteristics 17. ORS is not included in this list, as they do not provide clinical recommendations but assist providers in treatment and dispensing decisions, making them “expressly exempt from FDA regulation.” 18 Despite this exemption, ORS is deeply integrated into clinical workflows, shaping prescribing practices, influencing prescribing practices, and occasionally constraining patients’ access to treatments deemed essential 19. Its widespread influence, exemplified by initiatives like PDMPs, has significantly shaped healthcare decision-making, leading to a notable 40% reduction in prescribed opioid medications from 2006 to 2020 20. Acknowledging the potential discriminatory implications of ORS is imperative, as it could significantly impact pain patients’ access to medicines. Our study highlighted the observed inherent measurement bias and advocated for transparent and ethical AI use to ensure equitable access to pain relief and safeguard vulnerable patient populations.