A Methodology to Create Mother-Baby Dyads Using Data From the Pediatric Health Information System

In this retrospective cohort study, we linked maternal and newborn discharges using inpatient data from hospitals participating in PHIS (Children’s Hospital Association, Lenexa, KS), an administrative billing database that contains encounters from 49 tertiary care pediatric hospitals in the United States. Hospitals submit encounter-level data including patient demographic characteristics, diagnostic and procedural codes (International Classification of Diseases, 10^th Revision, Clinical Modification [ICD-10]), and detailed billing information including pharmacy, imaging, and clinical billing data. Data quality and reliability are assured through a joint effort between Children’s Hospital Association and participating member hospitals.

Among hospitals participating in PHIS, only a subset has large volumes of routine newborn deliveries and a smaller subset of those submit their maternity discharge data because PHIS is primarily a pediatric database. Among the 49 PHIS member hospitals, n = 18 had any babies born within that hospital and n = 25 had any submitted maternal discharges during the study period, whereas n = 15 hospitals had both and were considered for inclusion. We restricted this analysis to hospitals with at least 8000 submitted maternal delivery discharges and at least 8000 submitted newborn birth discharges over the entire study period to identify hospitals regularly submitting perinatal discharges for both mother and baby. This threshold was determined by examining the distribution of maternal delivery discharges and newborn birth discharges by hospital over the entire study period. The included PHIS hospitals are unique because they have large volumes of routine births, meaning they are not generalizable to all other PHIS hospitals. Per the PHIS Data Use Agreement and to maintain PHIS member hospitals’ confidentiality, we do not identify the hospitals included in the study.

We queried PHIS for singleton births for newborns discharged between January 1, 2016, and December 31, 2020, along with maternal delivery discharges for singleton deliveries during the same period. We restricted the sample to singleton births as multiples would have been excluded during the matching process. We excluded any newborns born elsewhere and transferred to the PHIS hospital on the day of birth using source of admission to restrict to those born in the PHIS hospital. Newborn inpatient discharges were identified by ICD-10 diagnosis code (Z38.00: single liveborn infant, delivered vaginally or Z38.01: single liveborn infant, delivered by cesarean) with age at admission of 0 days. Maternal singleton delivery inpatient discharges were identified by delivery ICD-10 procedure codes (Supplemental Table 2) with a corresponding ICD-10 diagnosis code specifying single liveborn infant as the outcome of birth (Z37.0: single live birth). Any maternal or newborn discharges with additional ICD-10 diagnosis codes for multiple births were excluded because these were indicative of a coding error with codes for both singleton and multiple births. Because PHIS database queries are discharge date- based, we excluded maternal deliveries and newborn births that occurred outside of the study period but where the mother or newborn was discharged during the study period.

We examined within-hospital year-over-year volume of births and deliveries submitted to PHIS and excluded years with low volumes (eg, hospitals that started or stopped submitting data into PHIS in a given year) or a mismatch between birth discharges and maternal delivery discharges (Supplemental Table 3). To identify mismatches, we assumed the volume of singleton births should be approximately equal to the volume of singleton maternal deliveries in a given year and flagged any with a discrepancy greater than ±15% for the ratio of births to deliveries. All excluded hospital years (n = 5) exhibited at least a ±200% difference between births and deliveries.

Our deterministic matching algorithm linked newborn birth discharges to maternal discharges within hospital by date of delivery/birth, mode of delivery, and ZIP code. We identified mode of delivery by ICD-10 procedure codes for maternal discharges (Supplemental Table 2) and ICD-10 diagnosis codes for infants (vaginal delivery: Z38.00, single liveborn infant, delivered vaginally; cesarean delivery: Z38.01, single liveborn infant, delivered by cesarean). Date of delivery was extracted from the date of delivery procedure for maternal discharges and date of birth for newborns. We excluded maternal and newborn discharges for whom we were unable to ascertain the mode of delivery. Discharges with an unknown ZIP code were also excluded. Before matching, maternal and newborn discharges were deduplicated within a hospital by date of delivery/birth, mode of delivery, and ZIP code, excluding all duplicates. Newborns were then matched to maternal discharges within hospital by date of birth/delivery, mode of delivery, and ZIP code forming mother-baby dyads, retaining only exact matches.

We describe the number of maternal and newborn discharges excluded at each step of the matching process. We report the overall match rate (percent) and range by hospital. Descriptive statistics were reported to describe clinical characteristics at birth, hospital use, and neonatal conditions and diagnoses during the birth hospitalization for the matched newborn cohort. These included recorded birth weight and gestational age, length of stay, NICU admission, mechanical ventilation use, in-hospital mortality, complex chronic condition¹⁶  (CCC) classification, and common neonatal diagnoses. CCC classification and neonatal diagnoses considered all diagnosis codes assigned to the newborn during the birth hospitalization. Data cleaning, matching, and all analyses were performed using SAS version 9.4 (SAS Institute, Inc, Cary, NC). The deidentified SAS code used for cleaning and matching data from PHIS can be found at https://osf.io/gms36/?view_only=bd09a5523d9b4de184a543bcc0a9d915).

Six of the 15 PHIS hospitals (40%) who submit maternal delivery and birth discharges met the volume criteria and were included in the study. The initial cohort at these 6 hospitals, before exclusions, had a total of 115 945 maternal delivery discharges and 133 103 newborn birth discharges during the study period. These hospitals represented nearly all regions across the United States (Southeast, Midwest, Southwest, West). The included hospitals were all either women’s and children’s hospitals, tertiary care children’s hospitals with a labor and delivery unit, or were adjacent to and affiliated with a larger health system that includes an adult hospital. Exclusion criteria and the number of maternal deliveries and newborn births removed at each step are presented in Fig 1.

We excluded 2% of all newborn births born elsewhere and transferred to the PHIS hospital. Fewer than 0.3% of all maternal deliveries and remaining newborn births occurred outside of the study period (eg, discharged from the hospital in 2016 but born in 2015) and were excluded. Based on our year-over-year volume analysis within each of the remaining hospitals, we excluded 4% of remaining maternal deliveries and 15% of remaining newborn births. Among the included hospital years (N = 25 years; Supplemental Table 3), the average ratio of births to deliveries was 0.99.

Our next 4 steps in the exclusion process sought to identify and remove records with errors or missing data involving variables used in the final matching algorithm. We excluded deliveries and births with (1) multiple births recorded as a result of a coding error, (2) unknown mode of delivery, (3) more than one mode of delivery recorded as a result of a coding error, and (4) unknown ZIP code. Each of these steps removed less than 0.2% of remaining maternal deliveries and newborn births. Of note, no maternal deliveries or newborn births were excluded because of an unknown mode of delivery.

In the final step before matching, we deduplicated deliveries and births within each hospital by the date of delivery/birth, mode of delivery, and ZIP code and removed all duplicates found. This step excluded 17% of remaining maternal deliveries and newborn births.

We identified a matching maternal discharge for 92.1% of newborn discharges (n = 84 593/91 809). Within-hospital match rates ranged from 87.4% to 93.9%. The final matched cohort contained 84 593 mother-baby dyads.

For the matched newborn cohort, we describe clinical characteristics at birth, hospital use, and neonatal conditions and diagnoses during birth hospitalization in Table 1. The majority of newborns had a normal birth weight (91.2%, n = 76 446) and were full term (61.2%, n = 45 586) or early term (27.4%, n = 20 398), whereas 8.8% (n = 7404) were low birth weight and 11.3% (n = 8431) were premature (<37 weeks). The median length of stay was 2.00 days (interquartile range 1.00) and 14.2% (n = 12 080) of newborns had an admission to the NICU. Among all newborns, 9.5% (n = 8023) were classified as having a CCC based on their birth hospitalization and 25.0% (n = 21 169) had a diagnosis of neonatal jaundice.

Using data from PHIS, we demonstrate the feasibility of linking maternal deliveries to newborn discharges forming mother-baby dyads with a high degree of success. Overall, we found a matching maternal discharge for more than 90% of births, with similar match rates across hospitals. This is one of the first studies, to our knowledge, that used data from PHIS to form mother-baby dyads by matching maternal and newborn discharges. Our final match rate is similar to other studies linking maternal and infant data from other sources.^5,6,17,18  The final matched cohort spans 5 years across 6 hospitals with nearly 85 000 mother-baby dyads. Given that the PHIS database contains demographic, ICD-9/ICD-10 diagnosis and procedure codes, as well as detailed billing codes that are unique to PHIS, this large and geographically diverse cohort of mother-baby dyads could be leveraged to study a wide variety of perinatal topics examining neonatal hospital utilization and birth outcomes that may be affected by maternal characteristics at the time of delivery.

This study has a number of limitations. First, we relied on administrative billing data that may not be complete or accurate with regard to comorbidities and sociodemographic factors. As described, we found records with both modes of delivery recorded despite our initial query for singleton births and deliveries. Next, to link newborns and mothers, we limited our sample to a unique subset of PHIS hospitals with large volumes of routine deliveries and submitting maternal delivery data. Included hospitals were all either women’s and children’s hospitals, tertiary care children’s hospitals with a labor and delivery unit, or were adjacent and affiliated with a larger health system that includes an adult hospital. Only 6 of the 15 PHIS hospitals had both babies born within that hospital and submitted maternal deliveries; thus, our sample is not representative of all birth hospitalizations in PHIS. Additionally, although all regions in the United States except the Northeast were represented, this sample is not representative of all PHIS hospitals and may not be generalizable to all births and deliveries in the United States, including those at adult tertiary care, community, or other hospitals. However, the proportions of newborns in our matched cohort that were low birth weight (<2500 g), preterm (<37 weeks), or admitted to the NICU are similar to recent national estimates and results from other neonatal studies.^13,19–21  The included hospitals in this sample may, though, be broadly similar to other large urban adult tertiary care hospitals with large volumes of births including a mix of routine births and those referred for specialty care. A similar matching algorithm could be applied to datasets outside of PHIS that are more broadly representative and population-based, such as The State Inpatient Databases available through the Healthcare Cost and Utilization Project (Agency for Healthcare Research and Quality, Rockville, MD), which represents inpatient discharges from community hospitals across 49 states. In such datasets, however, the proportion of newborn discharges matched to maternal deliveries would likely be lower than described in our study because of the inclusion of large adult birth centers. In institutions with larger volumes of deliveries, the number of within-hospital duplicates (same ZIP code, same day, and same mode of delivery) may be higher, resulting in a lower number of matched births and deliveries because of more newborns potentially being excluded during the deduplication process.

Next, to use a deterministic algorithm to match newborns to mothers, we removed all duplicate births and deliveries at the same hospital from the same ZIP code on the same day with the same mode of delivery to retain only exact matches. Deterministic matching offers high specificity because of the need for exact matches at the expense of sensitivity, which is higher in probabilistic matching. However, our findings are similar to previous studies describing great success in linking records with both deterministic and/or probabilistic methods.^2,18,22,23  In the context of deidentified administrative data in PHIS, there are no other identifiers that can be used to match maternal discharges to newborn births. However, other data sources may contain additional identifiers that are sufficiently unique allowing for a probabilistic phase of matching which may increase the number of matches. Additionally, we are aware that deduplication by ZIP code has the potential to exclude more maternal deliveries and births from urban ZIP codes because these would be expected to have a higher volume of same-day births with the same mode of delivery. To determine if the excluded duplicate records were more likely from urban ZIP codes, we examined the urban-rural distribution and found similar distributions before and after deduplication for maternal discharges from urban ZIP codes (96.9% vs 96.3%) and newborn births from urban ZIP codes (96.9% vs 96.4%). Although the overall match rate following deduplication was 92.1%, the match rate was 76.7% among the cohort with complete information on the matching characteristics before deduplication and 63.6% among all newborn birth discharges extracted (including those with missing data and coding errors). Importantly, we were unable to validate any matched dyads because our institution does not have routine deliveries and therefore was not included in our sample. However, validation of this methodology is possible for PHIS member hospitals who have large volumes of deliveries and submit maternal data and who are able to unblind and identify their own institution’s discharges. Validation of this methodology is an important next step and area of future work.

Despite these limitations, this proof-of-concept study describing a novel linkage methodology using data from PHIS may be of interest to pediatric researchers. PHIS is a large national database of pediatric hospital encounters in the United States available to member hospitals contributing data and paying for access to a robust administrative data source commonly used for health services research as well as benchmarking. PHIS is also a timely data source, with both clinical and detailed billing data lagging by merely a month, relative to state and national data sources, which often lag by years.^9,24  PHIS contains robust and standardized estimates of cost and daily detailed billing codes (eg, pharmacy, clinical, imaging, laboratory), which are unique to PHIS. Although only a subset of PHIS hospitals has routine deliveries and submit maternal data to PHIS, the included hospitals represented nearly every United States region. Though the resulting dataset may not be generalizable to all births and deliveries in the United States, the linkage methodology we describe offers a large, timely cohort of matched mother-baby dyads that could be used to examine hospital use and birth outcomes for an array of neonatal conditions likely to be affected by maternal diagnoses or treatments at delivery. Inferences from studies conducted with matched cohorts derived using this methodology in PHIS may be limited to the hospitals included in the sample.

Potential areas of study with this matched cohort may include associations between maternal pain management during delivery and birth outcomes, assessment of disparities in birth outcomes by maternal sociodemographic factors, examining hospital variation in maternal and neonatal utilization in the perinatal period, and impact of maternal medication assisted therapy (MAT) on neonatal treatment and utilization for newborns with neonatal opioid withdrawal syndrome, among others. In the latter example, in the context of administrative data, maternal receipt of MAT is only available via the maternal delivery hospital use data; the diagnosis codes available in the newborn’s hospital use data are not specific enough to indicate in utero exposure to maternal MAT. The most specific diagnosis code to indicate neonatal opioid withdrawal syndrome is neonatal withdrawal symptoms from maternal use of drugs of addiction (ICD-10 diagnosis code P96.1), which does not differentiate between therapeutic opioid exposure, as in the case of maternal MAT (eg, methadone), from in utero exposure to nonprescribed or illicit opioids. Linking the maternal deliveries and newborn birth discharges will allow researchers to explore the associations between maternal MAT (including choice of treatment) and infants’ pharmacologic treatment on birth utilization, which would only be possible once the maternal and newborn data are linked. Of interest for future studies using this methodology and similar to recent estimates,^13,19–21  8.8% of newborns in the matched cohort were low birth weight, 11.3% were preterm, and 14.2% were admitted to the NICU.

This methodology may allow PHIS member hospitals interested in neonatal research, including those without routine births and/or maternal discharges, to leverage their access to this robust data to explore questions that can only be answered using linked birth and delivery data. It is important for future studies applying this methodology to PHIS data to carefully examine volumes of submitted maternal discharges and newborn births to determine hospital-level inclusion criteria, as we describe for this study, because PHIS member hospitals’ submissions may vary over time.

The linkage methodology we describe could be applied to other administrative datasets in addition to PHIS. In the absence of ZIP code, which may be unavailable in other data sources for privacy and security reasons, other geographically linked variables (eg, demographic fields from Census material) may be sufficiently unique to be leveraged for linkage, particularly in combination. To demonstrate the feasibility of using other geographically linked variables as a proxy for ZIP code, we explored the use of median household income for matching in lieu of ZIP code in our PHIS cohort. We found overall and within-hospital match rates that were virtually identical to those using ZIP code (overall: 92.1%; range by hospital: 87.4%–93.9%). Although median household income is not a unique identifier theoretically, there was virtually no overlap within a hospital where multiple ZIP codes had the same median household income (n = 18 overlaps of 2534 unique ZIP codes, 0.7%). Therefore, median household income can provide a suitable alternative to uniquely identify geographic area for the purposes of matching within hospital. Match rates may be further improved by including additional small-area geographic variables, if available, allowing further granularity to uniquely identify geographic areas (eg, block or tract-level Census variables). Although PHIS does not contain any additional geographically linked fields that are suitable for matching, these may be applicable to other data sources.

We demonstrate a deterministic linkage methodology using a large administrative dataset to link maternal deliveries and newborn birth discharges to form mother-baby dyads. The ability to link maternal and newborn records in administrative data sources like PHIS allows researchers to examine associations between maternal demographics, social determinants of health, and clinical characteristics on neonatal hospital use and birth outcomes, which is only possible when data are linked.

None.