6B-01 - Adapting an electronic health record-based congenital malformation algorithm to insurance claims data lacking death information

Background: Accurate algorithms are needed to support real-world studies of medication safety in pregnancy. A validated electronic health record (EHR)-based algorithm developed by Kharbanda et al. is useful for identifying major structural defects. However, it is unclear whether this algorithm—which incorporates death information—can be adapted to claims data sources that lack death data.

Objectives: To modify an EHR-based congenital malformation algorithm for use with claims data and quantify the impact of missing death information on malformation prevalence (overall and by type).

Methods: Using the Merative™ MarketScan® Commercial Database (2006-2015), we created a linked cohort of birthing parents and their liveborn, singleton infants. We restricted to years with death available via inpatient discharge status; death was redacted after 2015. We identified malformations using diagnosis codes on infant claims ≤365 days after birth and birthing parent claims ≤30 days after birth. Separately, we used Kaiser Permanente Washington (KPWA) integrated EHR-claims data to create a cohort of liveborn infants (2007-2022). We identified malformations using diagnosis codes on infant claims in the first year of life and captured death via EHR, enrollment, and linked state death certificate data. In KPWA, we identified malformations with prevalence ≥1 per 1000 livebirths (n=7). We applied a validated EHR-based algorithm that defined malformations at the level of organ system; diagnosis codes were required on ≥2 dates for most malformations. We aligned 22 of 25 specific malformation definitions across ICD-9-CM and ICD-10-CM codes. Using MarketScan data, we calculated malformation prevalence using the algorithm with and without death information then compared results to published estimates. In MarketScan and KPWA data, we assessed the impact of death on malformation identification by calculating sensitivity (gold standard = definition with death data).

Results: Among 1,294,262 linked infants in MarketScan, malformation prevalence was 20.7 and 20.8 per 1000 livebirths using the algorithm without and with death, respectively. Prevalence estimates were generally comparable to the literature. In MarketScan, sensitivity ranged from 84% (21/25 prune belly) to 100% (eg, 2231/2232 pyloric stenosis); 17 of 22 malformations had sensitivity ≥95%. In 79,513 infants in KPWA data, sensitivity of the approach omitting death data ranged from 96% (160/167 severe cardiac defect) to 100% (eg, 65/65 pyloric stenosis).

Conclusions: We successfully adapted a highly accurate malformations algorithm from EHR to claims data. Omitting death information did not meaningfully impact sensitivity, suggesting this algorithm can feasibly be applied to data sources lacking death information.