In a world where electronic healthcare data are becoming increasingly used for the purposes of clinical trials and epidemiological research, there is a need for researchers to understand whether additional information can be gained by linking two (or indeed more) electronic health record data sources together. However, where there is overlap between the constituent data sets, such as with coding of clinical conditions, the researcher needs to decide which data set to rely on for identifying cases, or indeed whether combining information from both the data sets may be of value. Our study demonstrates that the method of coding MI, IHD and CVD appears to result in identification of different types of patient, in particular as characterised by prescribing and case fatality rates. Incident rates of disease also vary depending on the coding method used.
Previous work examining the epidemiology of cardiovascular disease has been conducted in Scotland using routine clinical data. Primary care data have been used to demonstrate that IHD is a common problem associated with male gender, increasing age and socioeconomic deprivation.12
Yet the recording of IHD data varies in general practice with different methods used for case detection.13
Furthermore, external factors such as payment-for-performance have been shown to improve the recording of IHD-related health indicators.14
Such incentivisation was introduced to UK general practice (but not hospital practice) in 2004, and so it is possible that this may have reduced the discrepancies between hospital and GP data in our study. Interestingly, pooling of GP and SMR records has previously been advocated for detecting MI cases,15
and pooled GP/SMR data from the same data set we used have demonstrated differences between cohorts of incident and prevalent MI.16
However, the effect of using only one component of such a data set has been hitherto unknown.
Reasons for differences in incidence rates and patient characteristics
Our data do not allow us to determine the exact cause of our findings, but a number of hypotheses may be proposed. Incident disease is reassuringly similar between GP and hospital groups for MI and CVD. The lower incidence of IHD for the GP group reflects the fact that many patients will have had relatively stable coronary disease for a number of years but not necessarily required acute hospital admission. Thus, many GP episodes of IHD do not count as true incident cases as they have had prior contact with the GP, whereas a higher number of hospital episodes are incident cases as these patients have never been previously admitted. The lower incidence rates for the paired GP/hospital group, and higher incidence rates for the pooled GP/hospital group, are inevitable consequences of the way in which the two data sets are united, although the magnitude of these differences will nonetheless reflect the degree of inconsistency in coding between the two. Furthermore, it would appear that because the paired GP/hospital data considerably underestimate the true disease incidence, it is probably not a useful method for identifying cases, even though such cases might be more rigorously identified. In addition, the increase in incidence rate using the pooled GP/hospital data demonstrates the potential advantage of combining two data sets, over use of a single data set, from the perspective of improving case finding.
The discrepancies in death rates are probably relatively straightforward to explain. Acute MI admission has a high case fatality,1
but those surviving beyond discharge have a much lower case fatality subsequently. It seems likely that the GP may fail to record the cause of death in patients who do not survive the hospital admission, thus resulting in the lower case fatality rates observed in the paired GP/hospital coding group. Furthermore, it is possible that patients coded only by the GP may represent ‘less serious’ illness, where hospitalisation is not deemed necessary by the GP. It is recognised that many patients suffering relatively minor strokes may not be admitted to hospital,17
resulting in lower case fatality for CVD in the GP group, although with the growing availability of active treatment options for ischaemic stroke in the form of thrombolysis, this may well change. We used national mortality data to identify deaths from both GP and SMR data sets, so discrepancies in recording of death between GP and hospital are unlikely to explain the differences in case fatality rates observed. Furthermore, the majority of paired events share exactly the same date, suggesting that retrospective date entry by the GP of the hospital event is common, and thus, there is no reason why this could not be carried out for fatal events.
The higher prescribing rates for IHD in the paired coding group are probably due to GPs responding appropriately to secondary care instigated intervention, reflected in appropriate treatment. That such differences were not observed for MI may be due to better communication and awareness for this specific condition compared with other IHD, such as angina, meaning that prescribing in the hospital group appears just as good as for the paired GP/hospital group. However, fewer MI events may have left us underpowered to detect differences. The lack of difference in the GP and paired groups for CVD may reflect poorer awareness of stroke management guidelines18
in comparison with coronary heart disease, and so prescribing rates are consequently no higher in the paired group. The lower prescribing rates of statins and antiplatelet agents in the CVD hospital group may reflect the GP being unaware of these patients' clinical need resulting in undertreatment; this is supported by the higher prescribing rates in the paired group. The differences in other patient characteristics—specifically smoking and comorbidity—are less easy to understand but may represent increased disease severity and mortality in hospitalised smokers and multimorbid patients. The small differences in age (<3 years) seem unlikely to be clinically relevant, although may be pertinent from the public health perspective. Finally, it may be that miscoding of diagnoses may explain some of the above differences; for instance, heart failure may be used as an alternative but incorrect code for MI.19
Furthermore, the introduction of sensitive troponin assays has influenced MI detection rates20
; it is possible that lack of familiarity among some clinicians for the resulting terms (eg, non-ST elevation MI, acute coronary syndrome) may result in inaccurate diagnoses being recorded.
This study has highlighted important issues related to patient coding and linked data, but although it has the advantage of using a reasonably large routine data set, linked at the individual patient level, a number of issues and limitations should be considered. The relatively small number of GP surgeries (40) may not have been fully representative. In addition, the number of events is relatively small, and given the conservative nature of the χ2
test, this increases the possibility of type 2 errors; thus, a larger data set may have identified more differences between groups. We restricted our analysis to four simple coding groups—GP, hospital, paired and pooled GP/hospital. However, it is clear that there are many further ways of categorising events, including the presence or absence of prior or subsequent coding based on the alternative half of the data set. For instance, an incident GP event with a historical hospital event may be coded differently to a GP event with no previous hospital record. However, we found that many of these theoretical categories have only a handful of cases. Furthermore, even when we examined six or seven separate smaller coding categories, similar differences in patient characteristics persisted between groups (data not shown). Our choice of four main groups was therefore a pragmatic one, which reflects the choice that would face a researcher dealing with a similar linked data set. The decision to use a 30-day limit for pairing data could also be questioned; we are unable to prove that these two events are truly the same clinical episode. The choice was again, therefore, partly pragmatic, although supported by examination of the distribution of time gaps between the GP and hospital data. We did not limit the lead-in time period prior to 1 January 2005 in any way. Length of GP records is generally greater and more variable than SMR records, and there is the potential to see a lower number of new incident events among persons with longer GP records. Our study used routine GP data, and it is possible that such profound differences may not be found with research-standard databases, such as General Practice Research Database (GPRD).21
Nonetheless, work linking primary care research databases to hospital (and other) records is ongoing, and the issues raised by our study must be acknowledged. The SMR data set only records hospital events in Scotland and thus fails to capture events in elsewhere in the UK or abroad. Similar issues face the English equivalent Hospital Episode Statistics, and a UK-wide hospital events data set would be valuable. SMR (and Hospital Episode Statistics) also provide multiple diagnostic codes for a single event. We elected to use all six diagnostic positions to ensure maximum capture of relevant hospital events. However, the robustness of low-priority diagnoses might be questioned. Nonetheless, we found similar results when we used only two diagnostic positions (data not shown). We also did not examine miscoding of events—for example, a code of angina being used rather than the code for MI. Coding of SMR is considered 99% complete and 88% accurate8
; corresponding metrics are not available for PTI data (although the completeness and accuracy of Read coding of morbidity in Scottish general practice has been shown previously to be greater than 91%22
). Furthermore, the two data sets use different coding systems, so completely reliable comparison is not possible. However, we used relatively broad definitions, and the Read code system is based on ICD. Nonetheless, we may in particular have missed some administrative Read codes, which might have enabled identification of additional cases in the GP group. Of course, ideally further validation of the coding should be conducted; linkage to laboratory data might be one way of achieving this. Finally, our 30-day limit for prescribing was selected from a pragmatic perspective. However, it is possible that patients who were admitted for over 30 days would not have had a new prescription issued by the GP within the 30-day post-event period, resulting in an apparent underestimation of prescribing. We believe that these numbers will be relatively small, however, and unlikely to alter the overall interpretation of our findings.
Research and policy implications
These results have significant implications for linked data; the drug management, disease severity and to some degree the patient characteristics vary depending on how the disease cohort is defined. They also have implications for the use of unlinked routine data—use of isolated primary or secondary care data may result in a biased selection of patients. This may affect patient recruitment as well as the validity and reliability of such information sources as secondary data in clinical trials, including clinical outcomes. It is similarly relevant to the public health environment. Using linked data allows one to have a more robust definition, by using pairs of GP and hospital codes only, but it is clear that the apparent incidence of a disease will be considerably lower. Alternatively, linked data enable a looser but more inclusive disease definition, using both GP and hospital data, but not relying on the coding occurring simultaneously. When using separate data from only one source, one needs to take into account that patient characteristics may not be representative of the wider population. It is difficult to recommend one coding approach over another, however, and the decision will need to be based on the specific question being posed.