Researchers have derived several algorithms to identify emergency department (ED) oral anticoagulant prescriptions for patients with atrial fibrillation (AF) in a large health dataset. The retrospective cohort study identified algorithms that, they say, “can be selected to optimise specificity, sensitivity, PPV [positive predictive value], or NPV [negative predictive value]”, “depending on project goals”.
Published as a research letter in JAMA Network Open, authors Clare L Atzema (Sunnybrook Health Sciences Centre and ICES, Toronto, Canada) and colleagues write: “The ability to identify these prescriptions in large datasets could facilitate future studies aimed at increasing such prescriptions and could support long-term monitoring of ED oral anticoagulant prescription patterns.”
AF is commonly seen in EDs, where initiating oral anticoagulation has been associated with higher long-term use than when prescribing occurs in outpatient care following discharge. As current datasets do not identify ED prescriptions, the researchers used the National Ambulatory Care Reporting System, and identified patients who had visited one of 20 Ontario EDs with a primary diagnosis of AF from 2009 to 2014. Patients were excluded if they would not benefit from oral anticoagulant agents, or had been taking them within 90 days, were <65 years (because there was no comprehensive outpatient medication data for them), or had been admitted to the hospital. Data were abstracted from patient medical records for the first visit per patient. Using unique encoded identifiers, the medical record data were linked to databases held at ICES.
The reference standard was medical record documentation of provision of an oral anticoagulant prescription by any physician (emergency or consultant). Atzema et al tested 32 algorithms. They varied the algorithms by:
- including oral anticoagulant prescription that were filled on the same day (day zero) as the patient left the ED, because some patients are sent to the ED by another clinician who may have prescribed the agent;
- counting the number of days following discharge that patients filled a prescription;
- excluding prescription fills that occurred after an outpatient visit with a relevant clinician, because that clinician may have written the prescription being filled;
- excluding patients with a history of venous thromboembolism, and
- removing patients who had seen a relevant clinician 30 days before the emergency visit.
Sensitivity, specificity, PPV, NPV, and 95% confidence intervals (Ci) were calculated. Of 2,015 qualifying patients, 65% had a CHADS2 (congestive heart failure, hypertension, age >75 years, diabetes, and stroke) score of two or higher. Emergency department physicians prescribed warfarin for a median of seven days and direct oral anticoagulant (DOAC) agents for a median of 30 days. Inclusion of day zero substantially improved sensitivity—for example, 34% in patients analysed in one algorithm where prescriptions were filled between days one and three, to 83% in patients analysed in a different algorithm where prescriptions were filled between days zero and three.
Using a longer follow-up period for prescription fills lowered specificity (from 91% to 74%, prescriptions filled between days zero and 30), and only slightly improved sensitivity (83% to 88%). After restricting to patients who filled a prescription before they saw another clinician, the number of false-positive results (that is, prescriptions filled that were not written in the ED) decreased, with specificity improving from 74% to 97%. However, sensitivity decreased from 88% to 68%.
Removal of patients who had an outpatient visit before the ED visit further increased specificity (99%) but decreased sensitivity (24%). Excluding patients with a prior venous thromboembolism resulted in minimal change. One algorithm maximised specificity (91%, 95% CI 90–92) while maintaining reasonable sensitivity (83%, 95% CI 79–87) (PPV 70%, 95% CI 66–74, NPV 96%, 95% CI 95–97). A prescription filled prior to seeing another relevant clinician provided high specificity (97%, 95% CI 96–98).
Among the limitations that the authors list are that medical record documentation of ED prescription provision was used as a reference standard, and that the results may not apply to patients <65 years.