Performance of an algorithm for identifying pregnancy outcomes in commercial health plan claims data

Background: There is increasing interest in using administrative data to examine pregnancy outcomes. The accuracy of using claims to identify particular pregnancy outcomes is not well known.

Objectives: The aim of this study was to assess the performance of an algorithm for identification of pregnancy outcomes within a commercial insurer's administrative database as compared with medical records.

Methods: In a retrospective study of pregnant women with psoriasis or chronic inflammatory arthritis and a general population comparator group, an 8.5% random sample of pregnancies, stratified by pregnancy outcome, was identified in a large claims database using systematic tracking of real kids, a process that identifies pregnancies and links mothers and babies in administrative claims. Outcomes for live births (single and multiple), non‐live outcomes (stillbirth, spontaneous, and non‐spontaneous abortions), and unknown outcomes were identified using a claims algorithm. Medical charts were sought and reviewed to confirm outcomes in claims. Positive predictive values (PPVs) and 95% confidence intervals (CIs) were calculated to estimate the proportion of claims that were true cases.

Results: Medical records were received for 300/457 pregnancies. Outcome data were recorded in the procured medical records for 180/232 live birth and 53/55 non‐live birth claims. The PPV for claims‐based live birth outcomes was 98.3% (95%CI: 94.8–99.6). All 53 charts for non‐live outcome claims were confirmed as non‐live [PPV = 100%, (95%CI: 91.6–100.0)]. Finer distinctions within these categories were also assessed. Among live births, the PPV for claims identifying single full‐term live births was 97.2% (95%CI: 92.6–99.1), but for multiple live births, it was 18.8% (95%CI: 5.0–46.3). Among non‐live births, the PPV for spontaneous abortions was 100% (95%CI: 86.0–100.0), but for non‐spontaneous abortions, it was 23.1% (95%CI: 6.1–54.0) with 10/13 of those claims described in charts as spontaneous.

Conclusions: Our algorithm performed well in discriminating live and non‐live pregnancy outcomes and in identifying spontaneous abortions but did not perform well for differentiating multiple live births or non‐spontaneous abortions.