Algorithmic opacity in opioid risk scoring and the need for transparent AI regulation

Area of Science:

• Health Informatics
• Pharmacovigilance
• Health Services Research

Background:

• NarxCare® is a proprietary opioid risk scoring system integrated into Prescription Drug Monitoring Programs (PDMPs).
• Patient complaints have been raised regarding the NarxCare® system.
• The accuracy and reliability of opioid risk assessment tools are critical for patient safety and appropriate prescribing.

Purpose of the Study:

• To independently evaluate the performance of the NarxCare® opioid risk scoring system.
• To assess the system's precision and F1 scores against established benchmarks.
• To determine if incorporating socioeconomic covariates improves the system's accuracy.

Main Methods:

• Reconstructed NarxCare® models using technical specifications.
• Utilized Prescription Drug Monitoring Program (PDMP) data.
• Incorporated IQVIA PharMetrics® Plus Closed Health Plan claims database.
• Included socioeconomic covariates in the analysis.

Main Results:

• Reconstructed models demonstrated precision values significantly below the benchmark (0.01-0.32 vs. 0.75).
• F1 scores were substantially lower than the benchmark (0.02-0.39 vs. 0.65).
• Performance remained poor even after adding socioeconomic covariates.

Conclusions:

• The NarxCare® system's performance, as reconstructed, falls considerably short of reported benchmarks.
• The system's low precision and F1 scores raise concerns about its reliability in opioid risk assessment.
• Further validation and potential improvements are needed for opioid risk stratification tools.