COMPASS leverages the standardized vocabulary made available through the Observational Medical Outcomes Partnership (OMOP). The Standard Vocabulary contains all of the code sets, terminologies, vocabularies, nomenclatures, lexicons, thesauri,
ontologies, taxonomies, classifications, abstractions, and other such data that are required for: 1) creating the transformed (i.e., standardized) data from the raw data sets; 2) searching and querying the transformed data, and browsing and navigating the hierarchies of classes and abstractions inherent in the transformed data; and 3) interpreting the meanings of the data211.
Within OMOP, the primary use of the vocabulary has been to translate source codes into standard concepts. For example, across the OMOP data community, conditions are coded using several different coding schemes, such as ICD9, SNOMED, MedDRA, Read, and OXMIS, but the vocabulary allows all sources to be standardized into a common vocabulary (either SNOMED or MedDRA). Similarly for drugs, many source capture prescriptions using NDC, GPI, VA Product codes, or Multilex, but these codes have been
mapped to RxNorm. The standard vocabulary also contains classification systems associated with its standards. For example, MedDRA provides a hierarchical structure of ‘is-a’ parent- child relationships whereby Preferred Terms (PT) are children of High Level Terms (HLT), which are children of High Level Group Terms (HLGT), which are children of System Organ Classes (SOC). The OMOP standard vocabulary offers several classifications for medical products. For example, RxNorm concepts are mapped into the National Drug File, Reference Terminology (NDF-RT), which provides classifications for mechanism of action,
physiological effect, chemical structure, and indication. Notably, RxNorm is also mapped to the Anatomical Therapeutic Chemical (ATC) classification maintained by the World Health Organization (WHO) Collaborating Centre for Drug Statistics Methodology, and the
National Drug Data File Plus (NDDF Plus) maintained by First DataBank. NDDF Plus provides multiple classifications for medical products, including FDA-approved indications, off-label uses, and contraindications. NDDF Plus is actively used in clinical design support tools for informing clinicians about medical information during prescription order entry. However, we are not aware of its prior use in population-level exploratory analysis of drug safety issues across observational healthcare databases.
COMPASS uses four attributes- therapeutic class, FDA-approved indications, off- label uses, and contraindications- as part of its automated heuristics, as shown in Figure 6. This graphic shows that ingredients can be mapped to each of these four attributes. It is worth highlighting that these attributes are actually mapped through RxNorm clinical drugs (which are concepts that uniquely identify product name and dose), but since active
without immediate exploration of dose effects, the attributes have been propagated up to the ingredient level.
Figure 6: Attributes of medical products used in COMPASS automated heuristics *Ingredient maps to these concept through RxNorm clinical drug, which contains product name and dose
An example of how these attributes are instantiated for a given medical product, lisinopril, is shown in Figure 7. All attributes have a many-to-many relationship with ingredients, meaning that each medical product can have one or more drug classes (here, lisinopril has only one, ACE inhibitors), one or more FDA-approved indications (lisinopril has three), one or more off-label uses (lisinopril has seven in total), and one or more contraindications (lisinopril has 40).
Figure 7: Example attributes for lisinopril
COMPASS uses these attributes to create automated heuristics for comparator selection, cohort restriction, and covariate adjustment. The logic for comparator selection is illustrated in Figure 8. The comparator group is initially defined by exposure to any medical products that have at least one of the same indications as the target drug of interest but don’t share a therapeutic class. To continue with lisinopril as a working example, the algorithm identifies all drugs that have an FDA-approved indication of either ‘hypertension’, ‘chronic heart failure’ or ‘myocardial infarction’. The drugs identified include ingredients from multiple drug classes, including: Angiotensin II Receptor Blockers (ARBs), such as losartan, valsartan, and candesartan; Beta Blockers, such as atenolol, metoprolol and acebutolol; Calcium Channel Blockers, such as amlodipine, nifedipine, and isradipine; diuretics, such as furosemide, amiloride, and hydrochlorthiazide; and other ACE inhibitors, such as enalapril, ramipril and captopril. The list is then restricted to those products who do not share a same therapeutic class. So, the other ACE inhibitors- enalapril, ramipril, captorpil- are removed from the indication drug list. Special consideration of combination products is taken to ensure ingredients that could be shared within the target drug are not erroneously included in
the comparator drug list. As such, hydrochlorothiazide is also removed from the comparator drug list for lisinopril because the combination of the two products is marketed (brand name: Zestoretic). The final list of comparator drugs reflect a set of alternative medicines that a patient could have been prescribed by a provider for at least one of the indications that the target drug is used for. Because most observational databases do not provide explicit patient- level information that ties diagnosis to prescriptions, pharmacoepidemiology studies
attempting to exploit the drug-indication relationship often do so by either assuming, inferring, or defining by explicit inclusion criteria. Moreover, pharmacoepidemiology studies commonly select a comparator drug for the unexposed cohort based on subjective assessment and clinical expertise. One reason for this approach is to minimize risk of immortal time bias that could be introduced if the unexposed population were defined by persons without any exposure (rather than an active alternative treatment). The COMPASS comparator selection heuristic provides an objective tool to construct a referent group to serve as the ‘unexposed’ population to compare with those exposed to the target drug of interest, and minimizes the potential bias introduced by subjective selection of only one ‘similar’ drug or class as an alternative treatment. The comparator selected varies by the drug under study as a proxy for ‘standard of care’ but does not reflect the notion of a ‘no treatment’ comparator group.
Figure 8: COMPASS automated comparator selection heuristic