(147) A Review of Algorithms Used to Identify People Who Use Nonmedical or Psychotropic Drugs in Electronic Healthcare Records and Claims Data

Publication Number: 478

Background: In both pharmacoepidemiology and clinical practice, various claims- and EHR-based algorithms are employed to identify people who use nonmedical and psychotropic drugs. However, the validity of these algorithms has not been evaluated in detail.

Objectives: To discuss a scoping review of nonmedical and psychotropic drug use-related algorithms applied to epidemiologic studies using claims- and EHR-based data.

Methods: We identified 5 validation studies of algorithms using diagnosis or procedures codes via structured PubMed searches. We included studies with a focus on nonmedical and psychotropic drug use, including substances such as opioids, stimulants, and sedatives. The Centers for Medicaid and Medicare Services Chronic Conditions Warehouse (CCW) definitions for drug use disorders were also reviewed. We summarize drug algorithm definitions, applications, and validation metrics, including the choice of comparators.

Results: We reviewed 5 validation studies and 3 CCW definitions. Of the 5 validation studies, study populations included people with documented opioid overdose (n=1), drug use-related infective endocarditis (n=2), and other general nonmedical drug use (n=2). The validation studies and CCW definitions employed a combination of codes for opioid use disorders, stimulant use disorders, sedative use disorders, drug overdoses, and hepatitis C virus infections. The CCW definition for “drug use disorders” also included individuals with nicotine use disorders, non-psychotropic substances (e.g., laxatives, steroids), and newborns being evaluated for prenatal substance exposure. Validation study metrics varied widely with positive predictive values ranging from 48% to 81%, to sensitivity ranging from 41% to 90%. Although limited to 5 studies, it appears that the validation metrics tend to differ by algorithm definition, population, and the choice of comparator to used as the “gold standard”.

Conclusions: The drug use-related algorithms we reviewed vary widely, both by definition and validity metrics. Reliance on diagnosis-dependency in algorithms may contribute to misclassification bias. The “gold standard” comparator employed in such studies may be unreliable as well. More validation studies of drug use-related algorithms should be conducted, and careful attention should be paid to the selection of the comparator group serving as the “gold standard” in these studies.