(A-281) Comparison of 4 algorithms to prospectively identify pregnancies in electronic health record data

Background: Multiple approaches exist to identify pregnancies in electronic health record (EHR) data, but it is unclear how algorithm specification affects cohort identification. This is particularly true for newer algorithm approaches that identify pregnancies prospectively—first identifying a pregnancy-related encounter and then an associated outcome, if available.

Objectives: To evaluate four algorithms for prospectively identifying pregnancies using EHR data from a single large health system in North Carolina (2014-2023).

Methods: We identified groups of encounters with at least one pregnancy outcome diagnosis/procedure code or at least one medication order for mifepristone or misoprostol at a dose specific to abortion care, within 30 days of each other. Outcome assignment differed based on the algorithm (A1, A2, A3, A4): A1 prioritized outcome type; A2 prioritized broadly categorized information type (procedures > diagnoses > medication orders) then outcome type; A3 and A4 prioritized more granular information type (delivery procedures > non-delivery procedures > diagnoses > medication orders) then outcome type. Outcome groups where the type of abortion was unclear were categorized as unspecified abortions (A3) or spontaneous abortions (A4).

Pregnancy episodes were identified by indexing patients at their first eligible prenatal encounter (i.e., an encounter with at least one diagnosis/procedure code associated with the prenatal period) and prospectively looking forward for an outcome, if available. Last menstrual period (LMP) was estimated using outcome dates and available codes. We compared the four cohorts (one cohort for each algorithm) by evaluating the overall proportion of outcome types and overlapping pregnancies within a person.

Results: The algorithms identified a total of 232,025-232,080 pregnancies, with 60.7% having an observed outcome. Across all algorithms, the LMP and outcome date matched for 99.6% to 99.9% of pregnancies. However, there was disagreement on outcome assignment: A1 had more stillbirths (897 versus 711-742 for the other algorithms) and more live births with twins or higher-order multiples (2029 versus 1485-1486), resulting in fewer uncategorized deliveries (4811 versus 5388-5529). A1 and A4 had more spontaneous abortions than A2 and A3 (13,270 and 13,663 versus 12,825 and 13,004, respectively), resulting in fewer abortions of an unspecified type (4153 and 4112 versus 4806 and 4815).

Conclusions: The four study algorithms similarly identified the presence of pregnancies within the EHR data. However, the choice of algorithm may have important impacts on the observed incidence of delivery and abortive outcomes.