Timely, high quality, clinically relevant obstetric research is important to improve patient care for women and their newborns, so that as often as possible new mothers leave hospital healthy and with a healthy newborn. Hospital admission data are an efficient and rich resource for conducting such research on conditions that affect pregnancy and its outcomes [1–4]. Hospital data are also vital to inform decisions about health planning and, in many places, hospital funding [5]. Obstetric data do not always capture the range of procedures and conditions that can affect pregnant women and their babies, or may capture them in a way that is more difficult to analyse, such as free text. In addition, population level obstetric data are less commonly available than hospital data. Hospital data are usually coded with diagnosis and procedure codes following international coding standards, and therefore provide a useful alternative or supplement to obstetric data for obstetric research [1–4, 6]. In order to perform effective population-based research, however, researchers need information on the extent to which the data accurately reflect the clinical situation.

Reporting accuracy of diagnoses and procedures in hospital data may be affected by changes in practice and as different conditions become the focus of guidelines, management and audits. In the case of gestational diabetes, for example, changes in thresholds for diagnosis following the publication of results of a large prospective blinded observational study [7] have resulted in rates more than doubling in New South Wales (NSW), from 5.7% in 2005 to 12.2% in 2014, and gestational diabetes now accounts for almost 30% of planned births before 39 weeks gestation (unpublished data). Hospital data, collected when a woman is admitted to a maternity facility (hospital or birth centre) during pregnancy and birth, are a valuable resource for examining such changes over time and their impacts on outcomes, however they are typically collected for administrative purposes such as billing, rather than for research. Validation is therefore important to assess the reliability of hospital data sources.

This study examines reporting of maternal medical conditions including diabetes, hypertension, thyroid conditions and morbid obesity (BMI>40kg/m²), which are associated with increased risks of adverse outcomes for mothers and babies [8–11], in hospital data. The most recent validation study assessing reporting of diabetes in coded NSW hospital data showed moderate reliability for gestational and high reliability for pre-existing diabetes, however it assessed data from 2002 [12], and predates the changes in diagnostic criteria. Previous studies have shown variable sensitivity for chronic hypertension (ranging from 44%–86%) [13, 14], gestational hypertension (10–71%) [15], thyroid conditions (10–97%) [15] and a systematic review [15] found only a single validation study for morbid obesity in hospital data, which showed low sensitivity (10%) [16]. Further, clinical practice and the obstetric population may have changed in the intervening years.

Validation studies may be conducted comparing coded hospital data with medical charts, however this process is time consuming and expensive to conduct, and tends to reflect only a short time period. An alternative approach is to compare reporting in two independent databases [17–19]. Here we compare reporting in coded hospital data, taken from the hospital records of birth and any pregnancy admissions, to obstetric data, collected from antenatal clinics and the obstetric record of the pregnancy and birth, using obstetric data as the reference standard.

The coded hospital data are drawn from the electronic medical record for a hospital admission, extracted by trained clinical coders following the International Classification of Diseases (ICD). A primary purpose of the coded diagnoses is to enable activity based funding (whereby hospitals are provided government funding largely on the basis of the treatment of patients), healthcare management and planning. The Australian Coding Standards provides guidelines to ensure consistency in clinical coding nationally. These limit conditions coded to those affecting patient management in the current admission, meaning that some chronic or other present conditions that do not affect the admission are not recoded. The guidelines also require coding to be substantiated by clear medical record documentation or confirmation from a clinician, and prohibit interpretation of results; for example, a recorded BMI of 40.1 kg/m² may not be assigned a code for obesity without an explicit, documented diagnosis of obesity in the notes [20, 21]. Coded data also inform various healthcare management and planning purposes, including the population-level New South Wales Admitted Patient Data Collection.

Obstetrics data were drawn from ObstetriX, a clinical database administered during pregnancy, birth and the early postnatal period, collected by midwives and partially self-reported. ObstetriX data inform the population-level New South Wales Perinatal Data Collection. ObstetriX may be considered an imperfect reference standard due to the different purposes and perspectives of the databases, however most reference standards are not without error and uncertainty [22, 23], and large population-level datasets have been demonstrated to be robust to random errors and omissions [24].

There were 38,343 singleton infants born at 24 weeks gestation and older at the two hospitals between January 2011 and December 2015 (Figure 1). Of these, 36,051 received antenatal care at their birth hospital and were included in the analysis. There were 35,928 infants born at 28 weeks and older and included in the analysis of gestational and pre-existing diabetes.

Sensitivity, specificity, positive and negative predictive values for the conditions of interest are provided in Table 1. Gestational diabetes was reliably reported over the period (sensitivity 83.6%, specificity 99.2%, PPV 92.7%), as was pre-existing diabetes (sensitivity 88.2%, specificity 99.9%, PPV 86.0%). Pre-eclampsia and eclampsia, gestational hypertension and any hypertension had good sensitivity (80.0%, 80.1%, 81.5%, respectively), but positive predictive value was moderate, showing that 30-40% of cases reported in the hospital data were not reported in ObstetriX (PPV 59.7%, 65.6%, 70.4%, respectively), and sensitivity and PPV for chronic hypertension were lower (sensitivity 53.5%, PPV 53.2%). Reporting of morbid obesity and thyroid conditions showed very low sensitivity (9.8%, 12.9% respectively), with moderate to high PPV (65.6% and 82.3%, respectively). Sensitivity increased for gestational diabetes, from 60.5% in 2011 to 95.7% in 2015 (Figure 2, Table 2), and increased less dramatically for pre-existing and any diabetes, gestational hypertension and any hypertension. Specificity and NPV were very high for all conditions examined.

Sensitivity and specificity were similar between the two hospitals (Supplementary Table 2). Specificity was very high for both hospitals. Sensitivity was slightly higher at Hospital One for non-diabetes related conditions, and there was less variation in sensitivity between hospitals for any hypertension compared to specific types of hypertension.

Consistency of reporting by characteristics of the pregnancy is provided in Table 2. For all conditions examined in detail, sensitivity was higher when a woman was nulliparous and where there was a hospital medical model of care (exclusive and with shared GP care), with the exception of gestational hypertension where midwife care had slightly higher sensitivity. Sensitivity to gestational diabetes was higher where the condition was managed with insulin or oral therapy compared to diet. Sensitivity increased by year for all conditions, most dramatically for gestational diabetes. Positive predictive value increased with time for pre-existing diabetes, but remained fairly stable for most other conditions, and decreased for hypertension. The median BMI of all those with a diagnosis in the hospital data of overweight, obesity or localised adiposity was 41.9 kg/m² (IQR 37.9–46.9 kg/m², n = 93).

The hospital data appeared to contain some misclassification of pre-existing conditions as conditions that arose in pregnancy (Table 2). Among cases of gestational diabetes that were recorded in ObstetriX but uncoded in the hospital data (667 of 4055, 16.4%), 12 (1.8%) were recorded as having pre-existing diabetes, while among cases of pre-existing diabetes that were recorded in ObstetriX but uncoded in the hospital data (36 of 306, 5.2%), 19 (52.8%) were recorded as having gestational diabetes (6% of total cases). Similarly, of the gestational hypertension cases uncoded in the hospital data (354 of 1778, 19.9%), 44 (12.4%) were reported with chronic hypertension, while 79 (54.1%) of chronic hypertension cases uncoded in the hospital data (146 of 314, 46.5%, or 25% of total cases identified in ObstetriX) were reported instead to have gestational hypertension.

Rates of the conditions investigated at the two tertiary hospitals were generally within what would be expected for the population (Table 3). However morbid obesity was lower than expected, while gestational hypertension and thyroid conditions were higher than expected for the population.

This study examined the reliability of coded hospital data for reporting of gestational diabetes and other maternal conditions, compared with ObstetriX data as the reference standard. Diabetes was well reported in the hospital data overall, with an increase over the time period from moderate to very high accuracy. Gestational hypertension, pre-eclampsia and eclampsia were moderately well reported, with high sensitivity but moderate PPV, while chronic hypertension had only moderate sensitivity and PPV. Using a broad ‘any hypertension’ category increased sensitivity and PPV and should be considered for studies using these data. Thyroid conditions and morbid obesity were poorly reported, with very low sensitivity and morbid obesity also showing moderate ‘false’ positives among cases that were reported, suggesting caution in the use of these data. Other sources of data on those two conditions should be sought where possible.

Sensitivity observed for pre-existing diabetes and gestational diabetes at the start of the study period were similar to that reported elsewhere [12, 21]. Sensitivity for chronic hypertension was within the range [13, 14] or slightly higher than reported elsewhere [21], while for gestational hypertension it was higher than previously observed [15, 21]. High sensitivity and low PPV for pre-eclampsia was also consistent with previous work [13, 26]. For thyroid conditions, sensitivity was at the lower limit of the range previously reported [15] and for morbid obesity sensitivity was consistent with poor ascertainment reported elsewhere [16]. The high specificity observed for all conditions was consistent with previous studies [12–16]. The higher sensitivity for more severe diabetes (among women with Type 1 or Type 2 requiring insulin compared to those not requiring insulin for pre-existing diabetes, and receiving insulin or oral therapy compared to those who did not for gestational diabetes), is consistent with previous research on diabetes [12] and findings that greater severity of a condition is associated with better reporting [17, 27].

Sensitivity increased over time for all conditions, although this was particularly stark for gestational diabetes and chronic hypertension. The general trend of increasing sensitivity over time may be partly related to the introduction of activity based funding, a government scheme for calculating hospital funding, which was ratified in 2011 and introduced in NSW in 2012. A systematic review of accuracy in UK hospital data found that a similar funding method, known as ‘results based funding’, was also associated with an increase in coding accuracy [28]. Sensitivity for gestational diabetes increased dramatically from 61% in 2011 to 96% in 2015, which is likely related to the change in diagnostic criteria and corresponding increased incidence and clinical focus on the condition and its management. This may have had flow on effects for pre-existing diabetes, for which the trend mirrored gestational diabetes.

Inconsistent reporting of thyroid conditions and morbid obesity are a reflection of the difficulties in capturing pre-existing, non-acute conditions from coded hospital data. Government policy and coding standards generally allow only conditions that affect the current admission to be coded, with the exception of diabetes and some other specific conditions. If a thyroid condition is well controlled, it requires negligible hospital resources during admission and is therefore unlikely to be recorded, although it may be recorded in the neonatal record since it can be more likely to impact the infant than mother. Morbid obesity may not be perceived as relevant to most hospital admissions during pregnancy, or staff may be reluctant to label a woman as obese due to negative connotations. The different timing of data collection between the two datasets may also play a role; for example, a person may report a history of thyroid disorder that is recorded in ObstetriX but has resolved or no longer requires medication, and therefore is uncoded in the hospital data.

We found evidence of misclassification of pre-existing conditions as conditions arising in pregnancy, with a tendency for chronic hypertension and pre-existing diabetes to be reported as gestational hypertension and diabetes, which has been previously observed [12]. This was more common for diabetes (25% of all cases) compared to hypertension (6% of all cases).

Rates reported in ObstetriX were similar to what would be expected for the population [29–34]. Slightly higher rates of gestational hypertension and thyroid conditions may reflect that the hospitals included are tertiary hospitals. Morbid obesity was lower than the expectation for the population, perhaps reflecting the fact that both hospitals are located in metropolitan Sydney, where morbid obesity rates tend to be slightly lower than rural and regional areas [35].

Consistency of reporting was similar for the two hospitals, suggesting that the results hold across different socioeconomic and ethnic patient populations, locations, and facilities. Slight differences in the rates may reflect the different demographic compositions, with Hospital One tending to have an older obstetric population with a different mix of ethnicities and comorbidities [36].

Our findings suggest that coded hospital data are a reliable source of information for gestational diabetes, pre-existing diabetes and all types of hypertension, with the exception of chronic hypertension. Chronic hypertension was reported moderately well, and reliability would be improved by using a grouped category for any hypertension. As thyroid conditions and morbid obesity were poorly reported, coded hospital data should be used with caution for these conditions and if possible, other sources of data should be sought. While there may be local idiosyncrasies in coding and reporting, NSW hospital data are coded following international coding standards, with ICD-10 widely used worldwide [15, 38], and previous studies have shown similar sensitivities between NSW hospital data and data from elsewhere [21]. We therefore consider these findings to be reasonably generalizable to other settings.