(B-211) Evaluation of a pregnancy algorithm using multiple data sources

Background: Limited information is available on algorithms that use multiple electronic health data sources to identify key features of pregnancy episodes.

Objectives: To evaluate the presence of conflicting pregnancy outcomes and accuracy of gestational age (GA) derived from a pregnancy algorithm.

Methods: We identified pregnancies (2017–2023) from electronic health records (EHR) and claims data from a Minnesota, US health system. We evaluated pregnancy outcomes (e.g.., livebirth [LB], spontaneous abortion [SAB], therapeutic abortion [TAB], stillbirth [SB]) by population: only a patient during pregnancy or delivery (patients), only enrolled in the health insurance plan (members), or both (member-patients). Pregnancies with conflicting outcomes were tabulated against the outcome with the highest value in the algorithm hierarchy. We assessed completeness of GA captured by the algorithm. For LB pregnancies with both delivery record and other GA sources (i.e., last menstrual period (LMP), obstetric records, infant records), we compared GA from each source to GA from the delivery record as the reference. For LB pregnancies linked to state birth records, GA with the highest value in the algorithm hierarchy (GA from delivery record, GA diagnoses based on claims, estimated date of delivery (EDD) from obstetric records) was compared to GA from the state birth record as the reference. To assess GA accuracy, we calculated the percent of pregnancies that had GA within 7 days of the reference and Pearson correlation (corr).

Results: We identified 196,817 pregnancies: 39% were patients, 34% were members, and 27% were member-patients. SABs were more frequently identified for patients, 18% vs15% for members or member-patients. TABs were more frequently identified for members, 6.7% vs 3.3% for member-patients or 1.1% for patients. Conflicting outcomes occurred in < 2% of pregnancies. SB was the most frequent conflicting outcome for LB, SAB for SB, TAB for SAB, and SAB for TAB. GA was available for 55% of non-LBs and 99.5% of LBs. When compared to GA from the delivery record, accuracy was low for LMP (81.7% corr: 0.83) and was high for infant record GA (99.8% corr: 1.0). When compared to state birth records, the algorithm-selected GA had high accuracy: GA from delivery record (99.4%, corr: 98.1), GA diagnoses based on claims (97.9% corr: 0.98), and EDD (96.2%, corr: 94.6).

Conclusions: Differences in pregnancy outcome distributions were observed across populations. Conflicting outcomes were rare. The GA selected by algorithm hierarchy had high accuracy in LBs. When evaluating non-LB pregnancy outcomes, studies may benefit from chart reviewing cases to better capture GA and resolve conflicting outcomes.