(A-288) What to Expect when You’re Expecting Missing Outcome Data: Development of Claims-Based Algorithms for Pregnancy Identification

Background: Existing claims-based pregnancy algorithms typically rely on observed end of pregnancy outcomes, such as livebirth, spontaneous abortion, stillbirth or other end of pregnancy outcomes. This has limited application in studies when low latency follow-up data are needed and looking at data occurring after a non-pregnancy related index date could introduce selection bias.

Objectives: To evaluate the performance characteristics of novel claims-based pregnancy algorithms that require only baseline data.

Methods: Using HealthVerity’s Inovalon Closed Claims Database, we selected females aged 12-55 years, with one year of continuous baseline medical and pharmacy enrollment, indexed on March 1, 2023. Pregnancy status at index date was determined based on International Classification of Diseases, 10th revision Clinical Modification (ICD-10-CM) codes denoting weeks or trimester of gestation. Codes indicating the number of weeks of pregnancy were prioritized over codes indicating trimester. If no pregnancy outcome was observed prior to or on the index date and the expected full-term delivery date occurred after the index date, the person was considered pregnant at index. In the first algorithm, weeks of gestation was evaluated using the code in the 365 days prior to index; if there were multiple encounters, the closest to index was used. In the second algorithm, the estimated weeks of gestation was refined if there was another diagnosis code in the 0-6 weeks prior such that the code denoting a greater gestational age was used. An existing, validated algorithm (Moll et al 2021) was used as the reference gold standard to evaluate the performance characteristics of our proposed algorithms.

Results: 5.1 million women met the study inclusion criteria. 93,442 (1.8%) were pregnant at index based on the first version of our algorithm, 92,976 (1.8%) based on the second version, and 108,993 (2.1%) based on the reference algorithm. The sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of our first proposed algorithm was 69.4%, 99.7%, 81.0%, and 99.3%, respectively. The sensitivity, specificity, PPV, and NPV of our second proposed algorithm was 69.3%, 99.7%, 81.2%, and 99.3%, respectively.

Conclusions: Our tested algorithms demonstrated modest and similar performance, without requiring follow-up data on pregnancy outcomes, with the first algorithm having reduced computational demands without sacrificing performance. The proposed claims-based algorithm may be useful for researchers conducting analyses in populations that may include pregnant women with follow-up constraints.