Development and validation of claims-based algorithms to ascertain gestational age in multiple gestations

Technical Summary: Development and Validation of Claims-Based Algorithms for Gestational Age in Multiple Gestations

Problem Statement
Accurate gestational age (GA) assignment is a cornerstone of pregnancy epidemiologic research, essential for classifying pregnancy timing, fetal/neonatal outcomes, and time-varying exposures. However, administrative claims databases, which are increasingly used for large-scale studies, do not directly record GA. While indirect indicators (diagnostic codes, procedure dates) are often used to estimate GA, existing algorithms were primarily developed and validated for singleton pregnancies. Multiple gestations (twins, triplets, etc.) exhibit distinct obstetric characteristics, including higher preterm birth rates and different healthcare utilization patterns, which may render singleton-derived algorithms inaccurate or inapplicable. Consequently, there is a lack of validated methods to estimate GA specifically for multiple gestations using claims data.

Methodology
The study utilized the Korean National Health Insurance Service (NHIS) database, covering 97% of the population, linked with the National Health Screening Program for Infants and Children (NHSPIC) to establish a reference standard.

• Study Population: The cohort included 26,714 multiple gestation pregnancies with delivery-related claims between 2015 and 2024. Pregnancies were excluded if the neonatal birth date fell outside the maternal hospitalization period or if the reference GA was outside the 24–40 week range.
• Reference Standard: Reference GA was derived from the NHSPIC linkage using the estimated date of confinement (EDC) and neonatal birth date, assuming a 280-day gestation duration.
• Algorithm Development: The dataset was split into development and test sets. Four distinct algorithms were developed:
  1. Algorithm 1 (ICD-10 Code-Based): Used preterm birth (PTB)-related diagnostic codes to assign representative GA values based on the median reference GA for pregnancies with and without PTB codes.
  2. Algorithm 2 (Ultrasonography-Based): Estimated GA using the timing of prenatal ultrasonography. Various windows were tested, with "first-trimester targeted sonography" selected as the optimal indicator based on development set performance.
  3. Algorithm 3 (Prenatal Screening-Based): Estimated GA using the timing of prenatal screening tests. The combination of Maternal Serum Alpha-Fetoprotein (MSAFP), unconjugated estriol (uE3), and inhibin A was selected as the best-performing indicator.
  4. Algorithm 4 (Hierarchical): A sequential framework that prioritized indicators by performance: Algorithm 2 (ultrasound) first, followed by Algorithm 3 (screening), and finally Algorithm 1 (diagnostic codes) if previous indicators were unavailable.
• Evaluation Metrics: Performance was measured using Mean Absolute Error (MAE), accuracy within ±7 and ±14 days of the reference GA, and the missing rate (proportion of pregnancies where GA could not be estimated).

Key Results

• Algorithm 1 (ICD-10): Demonstrated the lowest accuracy with an MAE of 10.3 days and 51.4% accuracy within ±7 days. It had no missing values.
• Algorithm 2 (Ultrasound): Achieved the highest estimation accuracy with an MAE of 5.59 days and 81.8% accuracy within ±7 days. However, it suffered from a high missing rate of 27.5% due to the lack of ultrasound claims for some pregnancies.
• Algorithm 3 (Screening): Showed strong performance with an MAE of 6.42 days and 77.4% accuracy within ±7 days, with a missing rate of 9.6%.
• Algorithm 4 (Hierarchical): Successfully achieved complete GA assignment (0% missing rate) by integrating the three approaches. It maintained high performance with an MAE of 6.14 days and 79.1% accuracy within ±7 days. The hierarchical approach assigned GA via ultrasound in 72.5% of cases, screening in 24.7%, and diagnostic codes in 2.9%.

Significance and Claims
The authors assert that this study provides the first validated claims-based algorithms specifically designed for multiple gestation pregnancies. The primary contribution is the demonstration that algorithms derived from singleton populations may introduce systematic misclassification when applied to multiple gestations, as evidenced by the shifted representative GA values observed in this cohort.

The study highlights that a hierarchical framework offers a pragmatic solution for administrative database research. By prioritizing the most accurate clinical markers (first-trimester targeted ultrasound) while falling back on screening tests and diagnostic codes, the proposed framework balances high estimation accuracy with data completeness. This approach eliminates the selection bias associated with excluding pregnancies that lack specific high-quality claims (e.g., ultrasound records), thereby supporting more robust epidemiologic investigations into maternal and neonatal outcomes in multiple gestations. The authors note that while the framework is interpretable and adaptable, external validation is required before application in other healthcare systems due to potential differences in prenatal care utilization and coding practices.