Discussion
To our knowledge, this is the first time that a systematic ontological approach has been used to identify pregnancies based on patient-level coded information. We believe that our system is capable of producing reliable results across different coding systems used within the same healthcare setting. Full details of the pregnancy ontology toolkit and its functionality can be found on our ontology webpage and online supplementary file 2.36 Our literature review identified examples of the multiple problems encountered in attempts to identify discrete pregnancies in patient-level data. There were many references to the poor quality of the data recorded about pregnancies, often worsened by a general lack of communication between the multiple agencies involved in the delivery of patient care.3 10 13–17 20 21 23 Apart from the Matcho study,16 we did not find any other ontologically based solution developed to overcome these problems.
The low median and mean numbers of codes processed by the toolkit for each detected pregnancy reflect the sparsity of pregnancy-related data available in the RCGP RSC database. The algorithm was functioning in most cases on no more than two or three codes and in many cases on just one code per detected pregnancy (figure 6) while having the flexibility to handle much larger numbers of codes when available. In many cases, there was no event coded in the data to represent either the pregnancy start date or the pregnancy end date so that one or both of these had to be calculated. This will clearly have had an adverse effect on the precision of these dates and thus also of the duration of the pregnancy. The paucity of the data also made it necessary to fine tune the parameter table to obtain plausible durations and also to ensure that search periods were optimised to avoid either misinterpreting late-coded entries as new pregnancies or alternatively incorrectly including them into a previous pregnancy when in fact they signalled a true new pregnancy.
Figure 6Number of coded entries processed per pregnancy in the database.
We started with a hierarchical (Read v2) terminology and our ontological approach enabled us to extend this into a polyhierarchical terminology (CTV3). In the coming year we will use this same approach to incorporate SNOMED CT37 38 concepts when UK GP systems have migrated to that coding scheme. Others could use our approach to work with ICD or one of its clinical modifications.
Age at start of any pregnancy ranged from 10 to 69 (table 2), but of the overall 405 591 pregnancies, more than 99% had ages in the range 15–44. The small numbers at the extremes were checked (RCGP RSC database review) and found to be genuine. There were differences relating to coding schemes and systems in use even allowing for the much smaller number of pregnancies relating to CTV3 use. Most notably, there were fewer non-term pregnancies in the CTV3 group than in the Read v2 group (figure 2). This cannot be explained by differences in the coding schemes because CTV3 provides the same or a greater range of concepts as Read v2. However, the system supporting CTV3 is significantly different from those supporting Read v2, and the difference may be due to a more stringent application of Information Governance preventing the extraction of sensitive codes relating to termination and abortion. In contrast, for pregnancies resulting in delivery per woman, we demonstrated that the algorithm performed consistently across the two coding schemes Read v2 and CTV3 (figure 2).
To improve accuracy, the Matcho study16 deliberately excluded any pregnancy represented by less than two coded entries. If the same exclusion had been be applied to our UK RCGP RSC data, we would have lost about 30% of all of the pregnancies detected. The CDM approach used can be expected to lose detail as clinical terms are condensed into CDM concepts. In contrast, our ontological approach enables us to leverage the richness of the clinical data and enable more inferences to be made at a more granular level. This may be particularly valuable where the data are sparse. Our approach may be preferable for those studies that need more reliable detection of any pregnancies represented in the data while the Matcho study CDM approach may be more appropriate for studies where precise duration of identified pregnancies is more important than reliability of pregnancy detection.
We demonstrated that a high proportion of the representation of each ontological concept was accounted for by a small number of subcategories suggesting that the task of mapping the pregnancy ontology to other coding schemes could be simplified by excluding the lower order subcategories (figure 1). Furthermore, this optimisation should enable the algorithm to scale well with respect to big datasets.
Our internal validation was limited to checking the working of the algorithm and reviewing the findings on the RCGP RSC database. There was no external validation such as checking original medical records or comparison with external data. Pregnancy data are available from secondary care. In England, these are called Hospital Episode Statistics,39 but they are only available 3 to 6 months in arrears. Such data were not available for validation of term pregnancies identified in this study due to time taken to obtain approval and cost. In any case, they are not without problems.40 41 However, such validation may be possible in future. The close agreement between ONS and adjusted RCGP RSC mean ages at time of pregnancy was reassuring.
The RCGP RSC network has a dashboard capability which enables us to feedback to practices and improve data quality. This pregnancy ontology will go live in our dashboard42 across the 2018/2019 season and report vaccine exposure to pregnant women alongside older people, high-risk groups and children which are fed back already. Other uses planned or in train include adherence to guidelines relating to drug use in pregnancy, associations between drug misuse during pregnancy and birth abnormalities, quality of postnatal care for women with gestational diabetes and pre-gestational type 2 diabetes.
