Feasibility of identifying important changes in care management resulting from cardiovascular magnetic resonance (CMR) using hospital episode data in patients who activate the primary percutaneous coronary intervention (PPCI) pathway

Patients were recruited from four NHS hospitals in England and Wales with 24/7 PPCI services (Bristol, Leeds, Swansea and Cardiff) between May 2013 and September 2014. Two hospitals (Bristol and Leeds) were defined as “CMR centres”, i.e. hospitals that had a local dedicated CMR service; the other two (Swansea and Cardiff) were defined as “non-CMR centres”, i.e. hospitals without access to a local dedicated CMR service. Patient identification, recruitment and consent have been described previously [9]. Patients entered the cohort at the point they activated the PPCI pathway (i.e. had an emergency angiogram), regardless of whether PPCI was carried out or not.

We used formal consensus (literature review and cardiologist expert opinion) to formulate consensus statements about important changes in management arising from the use of CMR in patients who activate the PPCI pathway. The formal consensus process has been described elsewhere [10]. Five patient subgroups were identified as potentially benefitting from CMR, namely patients who: i) have an out of hospital cardiac arrest (OHCA); ii) have “normal” (unobstructed) coronary arteries; iii) develop LV thrombus after STEMI; iv) have multivessel disease (MVD, defined as 2 or more vessels with > 50% stenosis pre-PCI); and v) underwent PPCI and have a suspected poor prognosis (assumed to be identifiable from CMR imaging biomarkers).

Three of the patient subgroups (patients who underwent PPCI, patients with unobstructed coronary arteries and patients with MVD) were defined using data collected at cohort entry (related to the index admission). Two subgroups, patients who had had an OHCA and those who went on to develop LV thrombus, could only be identified from HES / PEDW (data collected at cohort entry relating to the presence or absence of these conditions were missing for a large proportion of participants), using ICD-10 diagnosis codes I46 “Cardiac arrest” and I23.6 “Thrombosis of atrium, auricular appendage and ventricle as current complications following acute myocardial infarction”, respectively.

CMR exposure was defined as documentation of CMR in the imaging dataset (collected from “CMR centres”) within 10 weeks of their index admission; this time frame was chosen to capture both urgent and non-urgent (out-patient) CMR scans.

Initially, cardiologist members of the clinical team identified the key clinical “events” that they expected to reflect the important changes in management characterised through the consensus process. We then used HES and PEDW inpatient and outpatient data to ascertain these changes in management, compiling a list of diagnosis and procedure codes (International Classification of Disease, ICD-10; Office of Population Censuses and Surveys Classification of Interventions and Procedures, OPCS) representing the key clinical events with the help of a clinical coder. Most changes in management were defined by Boolean combinations of multiple ICD-10 / OPCS codes; for example, the “use of additional diagnostic tests during follow-up” was ascertained by OPCS codes for echocardiography (K58.5 OR U20.1 OR U20.2 OR U20.3 OR U20.4 OR U20.5), single photon emission computed tomography (SPECT) (U21.4); positron emission tomography (PET; U10.4 OR U21.3 OR U36.2), intra vascular ultrasound (IVUS; K51.2 OR L726), pressure-wire (K63.4 OR K63.5 OR K63.6) AND K51.8 AND Y44.2 AND Y53), radionuclide angiocardiography (U10.5) or computed tomography angiography (U10.2). We identified the key clinical events relating to the changes in management for each patient subgroup separately. The full code list used to identify changes in management is shown in the Additional file 1. For those patients who underwent CMR, any outpatient appointments that occurred between the index admission and the date of CMR were excluded. We did not use Accident and Emergency (A&E) data because most of the changes in management identified in the formal consensus process were deemed unlikely to have occurred in the A&E setting and diagnoses and procedures are not well coded in the A&E dataset.

Data were analysed using Stata/IC (V13, StataCorp LP, Texas, USA). The analysis was descriptive, with quantitative results expressed as counts and percentages. We calculated the frequencies of all events representing changes in management (e.g. number of new diagnoses, additional diagnostic tests, number / rate of outpatient appointments, etc.) in patients who did / did not receive CMR for each patient subgroup in the 12 months following the index admission.

We also calculated the time to first event (in participants in whom an event occurred) for each change in management (e.g. time to first new diagnosis, time to the first diagnostic test, time to next revascularisation in patients with multivessel disease, etc.) and summarised these with medians and interquartile ranges (IQR). Dates of events were available for all events, which allowed us to describe the time to first event for each change in management and for all subgroups, and to visualise the data using Kaplan Meier curves and other time to event plots. We did not compare differences between patients who did / did not receive CMR in any of the subgroups because this was not an objective; we were primarily interested in the feasibility of identifying relevant events and the frequencies of events in most subgroups were small. We did not attempt to derive a composite outcome because further research would be required to derive an appropriate weighting algorithm to combine the different clinical events.

The ICD-10 diagnosis codes and OPCS procedure codes of the clinical events representing important changes in management resulting from CMR are shown in Table 2. We successfully identified the events described in the table, except for medications; these are not available in HES/PEDW data. We were also unable to obtain data about medications on discharge for a large proportion of the cohort. Therefore, we could not identify changes in medication prescribed during follow up. All relevant non-ischaemic diagnoses were identifiable in admitted patient care data, as were procedures such as implantation of devices, repeat revascularisation and additional diagnostic tests. In contrast, we could not identify any of these from outpatient care data because ICD-10 diagnosis codes and OPCS procedure codes associated with each outpatient episode were missing for a large proportion of patients. All outpatient visits should have a diagnostic code associated with the visit. However, for at least one of their visits, 25% of patients had missing ICD-10 diagnosis codes, 74% of patients had missing OPCS procedure codes (which may indicate either that no procedure took place or that a procedure took place but was not recorded), and 74% of patients had unknown “Unknown and unspecified causes of morbidity” as their ICD-10 diagnosis code).

Tables 3 and 4 show the number of events and time to first event representing changes in management in the relevant patient subgroups identified in HES / PEDW up to 12 months after the index admission by CMR status and in CMR vs non CMR centres, respectively. It was possible to observe differences in both the frequency of events and time to first event by CMR status or in CMR vs non CMR centres. For example, across all subgroups, patients in CMR centres had more outpatient appointments and an earlier time to first appointment (see Fig. 1 for an example in the OHCA patient subgroup), and more frequent diagnostic tests and implantation of devices than patients in non CMR centres. Across most patient subgroups, patients in CMR centres and patients who had CMR seemed to have had an admission for an additional diagnostic test later than patients in non-CMR centres or patients who did not have CMR, respectively (see Fig. 2 for an example of CMR vs no CMR in the MVD patient group).

We have shown that it is feasible to identify clinical events expected to prevent future MACE and describe their frequencies and timing in relation to index PPCI pathway activation using only routinely collected NHS hospital episode data. We were able to identify relevant patient subgroups specified previously by combining data about the index event, which was obtained directly from hospitals, with admitted patient care (APC) hospital episode data. For each patient subgroup, we listed the “key clinical events” (e.g. new diagnoses or procedures, additional diagnostic tests, changes in medication, outpatient appointments) that reflected a change in management, identified through discussions with relevant clinicians. By looking at care pathways, where available, we confirmed that we captured all aspects of patient management in these key clinical events. We then compiled a list of the relevant ICD-10 diagnosis codes and OPCS procedure codes, identified these codes in hospital episode data for participants up to 12 months following the index procedure, and tabulated the events by patient subgroup.

We considered both the frequency of the events representing important changes in management and time to the first event for a composite primary outcome. This is because, for some changes in management (e.g. additional revascularisation for MVD patients or a new diagnosis in patients with unobstructed coronary arteries), it is the time to the event that represents the main change in management resulting from CMR. For other changes in management (e.g. additional diagnostic tests in patients with unobstructed coronaries), it is a change in the frequency/rate of an event this would represent the main change in management (e.g. a patient receiving fewer tests or outpatient appointments). We were able to tabulate the frequencies of events and their timing, and visualise these by CMR status (at the patient and centre level) for all patient subgroups. We were able to observe differences in the process of healthcare in all patient subgroups by CMR status. We did not explore possible reasons for apparent differences since this was not the aim of our study; the feasibility of implementing the consensus statements using hospital episode data was our main objective.

Composite endpoints combine several events of interest within a single outcome variable. A composite endpoint is intended to increase the frequency of a binary primary outcome and thereby increase the power of a study for a given relative treatment effect (e.g. risk, odds or hazard ratio). We did not attempt to combine clinical events to derive a composite primary outcome, either by subgroup or overall, mainly because we had no basis on which to weight different events. The derivation of a weighting algorithm to incorporate the “importance” of different clinical events would require further research, possibly involving a formal consensus process with clinicians, commissioners and patients, in order to provide a rationale for its use in a future registry assessing the clinical effectiveness of CMR in patients who activate the PPCI pathway. Another difficulty is that clinical events were characterised by different metrics (frequencies and rates) in our study, so some changes in management (e.g. the rate of cardiology outpatient appointments or additional diagnostic tests; changes in medication) cannot easily be included in a composite outcome for a time-to-event model.

Our study has several strengths. It is based on important changes in management and hospital activities previously identified systematically using formal consensus methodology, with input from a range of all stakeholders (clinicians, methodologists and HES coders). By using routinely collected hospital data, the study benefits from an efficient design and avoids the use of primary data collection. Our decision to use HES/PEDW data rather than local hospital systems was also based on the fact that follow up data would not be available from local hospital systems for patients who are repatriated to other hospitals.

A major limitation is that we had no data on medications, either at hospital discharge or during follow up. Medications prescribed at discharge should, in principle, have been easy to obtain from participating hospitals; however, none of the hospitals had electronic pharmacy records that could be linked by a common patient identifier. Medications (either newly initiated or changes to existing prescriptions) could not be obtained during follow up because there are no medication records in any of the hospital episode datasets. There was consensus that initiation, withdrawal or changes in medication represented an important change in management for all our patient subgroups [10], in particular patients with LV thrombus for whom anticoagulation therapy is the main change in management. This limitation should be overcome in the near future with the increasing implementation and use of hospital electronic prescribing (EP) systems [15]. A potential solution to the lack of medications data during follow up is linkage of data collected from hospitals and hospital episode data with data from the Clinical Practice Research Datalink (CPRD), a primary care dataset that has detailed information on medications (https://​www.​cprd.​com/​). Only a subset of GP practices (6%) contribute data to CPRD, so medications data would only be available for a proportion of patients in a registry.

Other changes in management (e.g. additional diagnostic tests) were likely underestimated because a large proportion of diagnostic tests are performed on an outpatient basis, and some tests, such as bedside echocardiography, are not recorded in the admitted patient care dataset. Although we could identify the frequency of outpatient appointments under the cardiology specialty (and could have done so for cardiac surgery or any other related specialty), we could not identify the tests themselves because test codes are not well recorded in outpatient datasets.

Our approach to constructing a proxy outcome reflects our original concept and appears to be feasible when the important changes in management are hypothesised to cause a change in the same direction in all events / activities included in a composite outcome for a subgroup. However, through the consensus process [10], it became apparent that some of the hypothesised changes in management arising from CMR would not necessarily result in a unidirectional change in the frequency or rate of an activity. In several instances, an important change in management was more appropriate targeting of a treatment activity to the patients who were most likely to benefit, without necessarily changing the overall frequency of the activity (potentially constrained in publicly-funded health services). For example, in patients with MVD, CMR may result in better (i.e. more cost-effective) targeting of additional revascularisation. In these situations, a simple evaluation of the frequency or rate of a proxy outcome would not detect any value of CMR to patients and the NHS. It is also worth noting that some hospital activities representing the important changes in management we identified (e.g. time to outpatient appointments and additional diagnostic tests) will be complicated by local NHS driven issues which may obscure any CMR test related differences.

Although the “concept” of creating a composite outcome reflecting changes in care management can be applied to any imaging technique for any disease condition, our composite outcome was developed from a consensus process with cardiologists which identified the changes in management resulting from CMR relevant to patients activating the PPCI pathway. The outcome cannot therefore be applied to other imaging techniques, even within the same population, since it relies on CMR being able to identify specific complications from the myocardial infarction (MI) that would not be identifiable by other imaging techniques (at least not in all relevant patient subgroups).