Database analysis part 1: RISK-II database

Database analysis part 1: methods

Data sources

We analysed patient records from the RISK-II database, which includes anonymised, linked, routinely collected data from (1) the Case Mix Programme (CMP) national clinical audit of adult intensive care,74 (2) Hospital Episode Statistics (HES) for England and (3) the Office for National Statistics (ONS) mortality database.

Case Mix Programme data are collected for the purpose of service evaluation and quality improvement in critical care.75 The CMP includes records for each admission to a participating adult high-dependency unit or ICU in England, Wales and Northern Ireland. Not all ICUs participated during the period for which data were extracted; coverage of adult general ICUs increased over the study period, reaching 100% in the final year of extracted data. Some, but not all, specialist ICUs participated (cardiothoracic ICUs were excluded from the analysis; see Inclusion and exclusion criteria). The CMP was used to identify the study sample, provide dates for the start and end of hospital admission and critical care, and to identify patient demographics.

The HES database is collected for the purpose of reimbursing NHS trusts for the provision of hospital services. The RISK-II database includes records from the admitted patient care section of HES, which contains one record for each ‘episode of care’ under one consultant during a hospital admission. One hospital admission may contain multiple episodes of care, one of which would generally correspond to the period in critical care, but there are differences between trusts in the way that these data are recorded; therefore, HES and CMP records do not align consistently. Each HES record includes up to 20 International Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10), diagnosis codes and up to 24 Office of Population Censuses and Surveys Classification of Surgical Operations and Procedures (OPCS-4) codes that were used to identify NOAF, comorbidities and diagnosis-specific rates of subsequent hospitalisation (see Identification of new-onset atrial fibrillation).

The ONS mortality database contains information about all of the deaths registered in the UK, and was used to derive indicators of mortality.

The anonymised RISK-II database is maintained by the Intensive Care National Audit and Research Centre (ICNARC) (London, UK) and was linked by NHS Digital (Leeds, UK) using a standard deterministic algorithm involving the NHS number (a unique patient identifier), date of birth, postcode and sex. The RISK-II database includes CMP records from 1 April 2009 to 31 March 2016, HES records from 1 April 2004 to 31 March 2016 and ONS records from 1 April 2009 to 31 October 2018.

Inclusion and exclusion criteria

We included the records of patients admitted to an ICU between 1 April 2009 and 31 March 2016 (5 financial years). Contiguous ICU admissions, representing transfers between ICUs and readmissions to an ICU within 1 calendar day, were combined into single records. We excluded admissions to cardiothoracic units, admissions lasting < 4 hours and patients aged < 16 years at the time of ICU admission.

Identification of new-onset atrial fibrillation

The codes used for identifying health conditions are summarised in Table 9. We defined patients as having NOAF during an ICU admission by identifying a CMP record that overlapped with a linked HES record that was the first HES record containing an ICD-10 code for AF for that patient anywhere in the database. Adopting a cautious approach to defining NOAF, we excluded patients with a linked HES record relating to the same hospital admission and contained an OPCS-4 procedure code for atrial ablation, pacemaker insertion or DCC during the same hospital admission from this group, because AF may have developed prior to ICU admission.

There are many possible ways that CMP records and HES records can overlap: the HES record may commence on the same day or on an earlier/later date than the CMP record, and similarly may finish on the same day or on an earlier/later date. Many of these scenarios result in uncertainty about whether the AF developed prior to ICU admission, in an ICU or after discharge from an ICU. We, therefore, implemented the following rules for classifying NOAF (Table 10):

• If the HES record commenced on the same day as the CMP record, then it was considered NOAF only if there was also a prior HES record from the same hospital admission (i.e. if the patient was admitted straight to an ICU from an emergency room and, therefore, had no prior HES record from that hospital admission then we assumed that the AF developed prior to ICU admission).
• If the HES record commenced on an earlier date than the CMP record, then the AF was assumed to have developed prior to ICU admission and not to represent NOAF.
• If the HES record commenced after the CMP record commenced and finished before the CMP record finished, then the admission was considered to represent NOAF.

If the HES record continued beyond the CMP record, then it was considered NOAF only if the discrepancy was only 1 day (to allow for minor discrepancies in data entry between the systems), otherwise it was assumed that the AF developed after discharge from the ICU. As a sensitivity analysis, we allowed for any amount of discrepancy in end date provided that other rules were met. This resulted in two possible definitions of NOAF: one for use in the primary analysis and an alternative approach described above for use in the sensitivity analysis. Both, however, reflect a cautious approach to identification of NOAF and exclude many scenarios for which we cannot distinguish NOAF from prior or subsequent AF. This has implications for the interpretation of results that will be highlighted below.

Identification of comorbidities and outcomes

Comorbidities were identified from any previously linked HES record that contained an ICD-10 diagnosis code for diabetes mellitus, hypertension, thromboembolism, valvular heart disease, dilating cardiomyopathy, pulmonary hypertension or heart failure (see Table 9). The date and cause of death were obtained from the linked ONS records. Subsequent hospital admissions were identified by linked HES records and classified as involving AF, stroke or heart failure.

Selection of matched comparators

Observational research using routine data has a fixed ‘observation window’ within which the data are collected (and within the RISK-II database, this window varies across the contributing data sets, as described in Data sources). Patients admitted to an ICU at different points in time are, therefore, followed up for different amounts of time. It is also common for the quality of routine data to vary over time and between contributors to a data set (e.g. between hospitals). To ensure that follow-up and data quality are comparable between patients with NOAF and any comparison group, we selected a cohort of comparators matched on hospital and month/year of admission to ICU.

Comparator patients were selected from all available admissions that were classified as neither NOAF nor prior AF (but including admissions for patients who subsequently developed AF). To ensure that patients with multiple admissions were not over-represented among comparators, one admission was selected at random from each patient’s set of candidate comparator admissions for consideration in the matching process. Matching was then performed with the largest ratio that could be supported while ensuring that at least 99% of patients with NOAF were included in comparisons.

Statistical analysis

Patient characteristics and comorbidities are reported as median with interquartile range (IQR) or as counts and percentages. To account for varying duration of follow-up of patients admitted at different points in time, outcomes were estimated using time-to-event methods with censoring of patients at the end of the relevant data set’s observation window (31 October 2020 for mortality or 30 March 2020 for hospitalisation). Mortality was estimated using the Kaplan–Meier cumulative incidence function from the date of ICU admission and, separately, from the date of hospital discharge among hospital survivors. The cumulative incidences of subsequent hospitalisation with AF, stroke and heart failure were estimated using non-parametric methods to account for the competing risk of death.76,77 Mortality was assessed at hospital discharge, and at 90 days, 1 year, 3 years and 5 years after hospital discharge, among hospital survivors. The cumulative incidences of hospitalisation with AF, stroke and heart failure were assessed at 1, 3 and 5 years after hospital discharge, among hospital survivors.

We estimated the associations between NOAF and outcomes before and after adjustment for patient characteristics and comorbidities using multivariable regression models adjusting for age, sex and comorbidities. Odds ratios (ORs) for hospital mortality were estimated using logistic regression. Hazard ratios (HRs) for mortality after hospital discharge were estimated using Cox proportional hazard regression. For subsequent hospitalisation with AF, stroke and heart failure, we estimated unadjusted and adjusted cause-specific hazard ratios (CHRs),78,79 censoring patients at death or the limit of follow-up. The proportional hazard assumption was tested by visual inspection of Schoenfeld residual plots. All covariates were modelled using dummy variables except for age, which was modelled continuously using a restricted cubic spline. Knot positions for the restricted cubic spline were selected in in accordance with the recommendations of Harrell.80

Finally, results for the primary analysis were compared with a sensitivity analysis that employed an alternative operational definition for NOAF, as detailed in Identification of new-onset atrial fibrillation.

Database analysis part 1: results

Data linkage and matching

The selection of records is summarised in Figure 2. Between 1 April 2009 and 31 March 2016, there were 965,576 admissions to 248 ICUs participating in the CMP, with links available to HES and ONS. After combining multiple records representing transfers and readmissions, a total of 919,801 distinct ICU admissions were extracted. Of these, 841,005 ICU admissions met the inclusion criteria. Of 8203 records identified as NOAF or possible NOAF, 8145 were matched to 48,870 comparators. A total of 4615 (56.7%) patients with NOAF and 27,690 matched comparators were included in the primary analysis.

FIGURE 2 Record selection and matching. (Figure not reproduced here; see the original article.)

How common is new-onset atrial fibrillation in critical care?

Of the 841,005 critical care admissions examined, 4615 (0.6%) admissions had a linked HES record indicating likely NOAF. A further 3548 (0.4%) admissions had a linked HES record indicating possible NOAF but where the HES record continued for > 1 day beyond the CMP record (the latter were included in the sensitivity analysis). Although the prevalence of NOAF using either definition appeared stable over time, the prevalence of prior AF (n = 165,150, 19.6%) increased over the first 5 years of the observation window (Figure 3). A reduction in prior AF in the final year of the data partly reflects the unavailability of procedure codes for identifying atrial ablation and pacemaker insertion in that year (a structural limitation of the database, which did not appear to affect the identification of NOAF).

FIGURE 3 Prevalence of AF in the RISK-II database. (Figure not reproduced here; see the original article.)

What are the typical characteristics of patients with new-onset atrial fibrillation in critical care and how to do these compare with other patients in critical care?

Patient characteristics and comorbidities are summarised in Table 11. Patients with NOAF tended to be older and have higher levels of comorbidity, especially hypertension, heart failure and valvular heart disease, than comparator patients without NOAF.

What are the outcomes for patients with new-onset atrial fibrillation in critical care and how do these compare with those for other patients in critical care?

The outcomes are summarised in Table 12 and are illustrated in Figures 4 and 5. Patients with NOAF were more likely to die, both during their hospital admission and after discharge, than comparator patients without NOAF. They were also more likely to be subsequently admitted to hospital with AF, stroke or heart failure.

FIGURE 4 Cumulative incidence of mortality from ICU admission, estimated using the Kaplan–Meier method. (Figure not reproduced here; see the original article.)

FIGURE 5 Cumulative incidence of mortality and hospitalisation after hospital discharge. (a) Mortality after discharge; (b) hospitalisation with AF; (c) hospitalisation with stroke; (d) hospitalisation with heart failure. Cumulative incidence of mortality estimated using the Kaplan–Meier approach. The cumulative incidences of hospital admission with AF, stroke and heart failure estimated using non-parametric methods to account for competing risk of death. (Figure not reproduced here; see the original article.)

How much of the difference in outcomes is explained by differences in patient characteristics and comorbidities?

Adjusted outcomes are summarised in Table 13, with model coefficients provided in Appendix 5, Tables 20 and 21. The excess risk of hospital mortality reduced by about half when controlling for differences in patient characteristics and comorbidities (OR 3.22, 95% CI 3.02 to 3.44, before adjustment, reducing to OR 2.32, 95% CI 2.16 to 2.48, after adjustment).

For outcomes post discharge from hospital, an examination of the Schoenfeld residuals indicated that the proportion hazards assumption was unlikely to be met for mortality or subsequent hospitalisation with AF, but was met for subsequent hospitalisation with stroke and with heart failure. In response to this, for mortality we partitioned follow-up into three time periods: 1–90 days, 90 days to 1 year and > 1 year. We then fitted separate Cox regression models for each time period (see Table 12). Results suggested that a similar proportion of the excess risk of death in the first 90 days after hospital discharge was explained by differences in patient characteristics and comorbidities as for death in hospital. After 90 days, adjustment for these factors explained all of the excess risk of death associated with NOAF (CHR ≈ 1.00, after adjustment).

The analysis of subsequent hospitalisations was complicated by the need to account for both the proportion hazards assumption and the competing risk of death. Because subsequent hospitalisation with AF exhibited the largest between-group difference, we elected to ignore the possible violation of the proportional hazards assumption and present analysis of the entire follow-up period, in keeping with the hospitalisation with stroke and heart failure. The results for hospitalisation with AF should, therefore, be interpreted as an average over the follow-up period that should not be assumed to be constant. Adjustment for patient characteristics and comorbidities indicated that about half of the excess risk of subsequent hospitalisation with each of AF, stroke and heart failure was explained by these factors.

Sensitivity analysis

For the sensitivity analysis, patients who had less-certain evidence indicating possible NOAF and their corresponding comparators were included in the analysis (n = 8145 patients with NOAF or possible NOAF and n = 48,870 comparators). Patient characteristics and comorbidities were similar between the primary and the sensitivity analysis (see Appendix 5, Table 22). However, hospital mortality fell from 43.9% among patients with NOAF in the primary analysis to 34.5% among patients using the expanded definition of NOAF in the sensitivity analysis (mortality among comparators was equivalent between the analyses) (see Appendix 5, Table 23). The results from regression models (see Appendix 5, Table 24) were consistent with the primary analysis in terms of the proportion explained by patient characteristics and comorbidities. There remained a small but statistically significant impact of NOAF on mortality > 1 year after hospital discharge; however, the CI overlapped with the equivalent interval from the primary analysis. Sensitivity analysis outcomes are illustrated in Appendix 5, Figures 11 and 12.