2.3. Medicare claims-based variables

We created Medicare claims-based variables to describe our cohort and to construct our algorithms. First, we obtained basic demographic information on the age and sex of the sample from claims. Next, we assembled lists of relevant Medicare claims for epilepsy diagnosis, stroke diagnosis, procedures for EEG and brain imaging, place and provider services, and medications (Table B.2). We were blinded to both the patients’ strata and the EHR based diagnoses in this step.

We used clinical knowledge to create four claims-based candidate variables for use in our algorithm: 1) epilepsy diagnosis; 2) stroke diagnosis; 3) EEG; and 4) specific AED (Table 1). “Epilepsy diagnosis” sums the number codes indicative of epilepsy as either a primary, secondary, or tertiary diagnosis. “Stroke diagnosis” sums the number of cerebrovascular codes indicative of stroke. “EEG” additionally sums EEG procedure codes coinciding with an epilepsy code. Lastly, “specific AED” accounts for the presence of claims of either levetiracetam, lamotrigine, phenytoin, valproic acid, or carbamazepine (the most specific AEDs prescribed for patients with epilepsy).

Variables constructed from claims

AED, antiepileptic drug; EEG, electroencephalography.

Description of claims-based variables.

We then reconstructed the initial sample using inverse probability sampling weights (IPW) and analyzed the claims-based clinical and treatment characteristics of the initial sample (i.e., we weighted our sample of 1,534 patients to represent our initial sample of 20,943 patients; Table 2). We presented the distribution of our claims-based variables (Tables B.3 and B.4) and our reference-standard EHR-based PSE and epilepsy diagnoses (Table B.5).

Reconstructed Sample Characteristics

Breakdown of reconstructed sample.