This analysis was a sub-study of the WAMACH study which used linked administrative data from the Hospital Morbidity Data Collection (in-patients) and Death Register from the Western Australian Data Linkage System, linked to PBS dispensing records.

The WAMACH study cohort has been described in detail previously and includes patients hospitalised in Western Australia for ischaemic heart disease (IHD), heart failure (HF) or atrial fibrillation [¹⁰]. The current study cohort was restricted to age 65 years or more, as seniors gain concessional access to PBS listed drugs. The remaining cost after the lower co-payment is met by the Australian Government, giving rise to a dispensing record in the PBS data. This captures the majority of warfarin users as approximately 90% of Australians aged 65–75 years, and 95% of those aged 75 years and over, had a concession card in 2004–2005 [¹¹].

The NVAF study cohort comprised all people aged 65-89 years admitted to a public or private hospital in WA from 1 July 2003 to 31 Dec 2008, with a principal or secondary diagnosis of atrial fibrillation/atrial flutter identified by the International Classification of Diseases (ICD) 10th revision Australian Modification (ICD-10-AM) code I48. Patients with comorbid cardiac valve disease, or with a prosthetic valve, were excluded, as were those undergoing long-term renal dialysis. Cohort entry was the first admission in which AF was recorded during the study period (the index admission). NVAF patients with an additional clinical indication for short (weeks) or longer term (12 months) OAC, i.e. joint replacement or for venous thromboembolism, were not excluded as the major aim is to assess long-term persistence/adherence to OAC in patients with a strong clinical indication (AF and high CHA2DS2-VASc score) and no contraindication, or reason for discontinuation such as major bleeding.

Linked data available for the AF cohort were all hospital admissions from 1980 to end of 2013, deaths (and cause of death) to end of 2013, and PBS medication records for the period 1 July 2002 to 30 June 2011. There was a minimum 12 months of PBS records before and after the index admission.

As observational cohort studies of therapeutics are at risk of immortal person-time bias [¹²], we used a landmark analysis method on patients alive at one year post-index admission (1-year landmark date).

We used the standard dispensing (supply) of 50 tablets, of any strength of warfarin, on at least one occasion in a one-year period as indicating exposure to warfarin during that period. Warfarin use at the landmark time point was identified according to supply in both the year prior to the index admission, and in the year after hospital discharge to the 1-year landmark. Warfarin exposure was thereby categorised as ‘persistent’ for patients with supply in both one-year periods, ‘discontinued’ for those with supply prior to index but not in the one year after, ‘initiated’ for those with no supply before index but with supply in the one year after, and ‘no warfarin’ if no supply in either period.

Age at the landmark date was grouped as 65-74, 75-79 and ≥80 years, approximating the cohort tertiles. AF at index admission was classified as either a ‘principal’ or a ‘secondary diagnosis’.

Comorbidities (heart failure, diabetes mellitus, hypertension, stroke/transient ischaemic attack/thromboembolism and vascular disease) identified from administrative hospital records in the prior 10-years were used to calculate the CHA2DS2-VASc risk score [¹³] at the 1-year landmark. A history of major bleeding in the principal diagnosis was identified, based on standardised bleeding definitions previously published [¹⁴,¹⁵], including intracranial haemorrhage (ICH) which incorporates intracerebral bleeding, subarachnoid haemorrhage, subdural and epidural bleeding, gastro-intestinal (GI) bleeding, and anaemia. Chronic kidney disease (CKD) was also identified.

Three outcomes were identified: time to death from any cause (censored at 3 years); time to a first stroke/TE (fatal or non-fatal) (censored at non-stroke death or 3 years); and time to a first serious bleeding episode requiring hospital admission, including fatal and non-fatal ICH (censored at death or 3 years). Outcome events based on hospital admissions were restricted to principal diagnosis only. The person-time incidence rate of stroke/TE and bleeding per 100 years (/100 PY) were calculated for each year from the landmark date.

The composite endpoint of stroke/TE was defined as ICD-10-AM codes I63 (ischaemic stroke), I64 (stroke, not specified as ischaemic or haemorrhagic) and I74 (arterial embolism and thrombosis) to identify fatal and non-fatal events. Combining I63 and I64 is recommended for identifying ischaemic stroke from electronic records for epidemiological studies [¹⁶], and is used in other population studies [¹⁷]. However, as nearly 90% of strokes as the ‘cause of death’ were coded as I64, a sensitivity analysis was done to determine the effect of excluding unspecified fatal and non-fatal stroke. All ICD codes are listed in Supplementary [Table 1].

Patient characteristics across warfarin exposure groups are described and compared using chi-square tests for categorical variables and ANOVA for quantitative variables. The annual sum, median, and interquartile range (IQR) of warfarin supplied (in milligrams, mg) were calculated.

Kaplan-Meier survival analysis was used to assess the associations between the outcomes and each categorical covariate, and the log-rank test used as a measure of their association with outcomes. Multivariable Cox regression models were used to compare outcomes across warfarin groups, adjusted for CHA2DS2-VASc score and other covariates (see [Table 1]).

A 2-sided p<0.05 was considered statistically significant. Estimated hazard ratios with a 95% confidence interval that does not include the value 1.0 are significant at the 5% level. SPSS Statistics for Windows (IBM Version 22.0. Armonk, NY) was used for all analyses.

The study was approved by the Human Research Ethics Committee (HREC) of the Department of Health (Western Australian 2014/11 and Federal Government) and the University of Western Australia (RA/4/1/8065), in addition to HREC approvals from all participating hospitals.

The datasets for this study, generated by the Data Linkage Branch of the Health Department of Western Australia are held at the University of Western Australia. The data are not publicly available due to confidentiality agreements between the providers and the study investigators.

Of the 14,634 alive at the one-year landmark 1,384 (9.5%) died in the following year; crude all-cause mortality was 9.0%, and 8.7% in the subsequent years.

Among the 3,631 deaths at three years from the one-year landmark date, the cause was coded as ‘vascular disease’ for 1,519 patients (41.8%) which included IHD (29.4%), ischaemic stroke (6.2%), and ICH (1.5%). Cancer deaths were 25%.

Unadjusted survival was greater among patients persisting with, or initiating, warfarin prior to the landmark date compared with those not exposed to warfarin, and poorer for those who had discontinued (Log Rank 83.35, df 3, p<0.001) [Figure 1]. Annualised survival, overall and according to warfarin exposure, is shown in [Table 2].

Figure 1. Kaplan-Meier survival to three years from the 1-year landmark after admission for AF (principal or secondary diagnosis) by warfarin use at the landmark date

---

---

Multivariate associations with survival to three years from the one-year landmark

Warfarin exposure group, a CHA2DS2-VASc score ≥3, age, male sex, comorbidity and previous events were all independently associated with 3-year all-cause mortality in 1-year survivors of AF [Table 3].

Stroke/TE was the principal diagnosis at the index admission for 727 patients (4.2%) among whom 38 (5.2%) died within 30-days. There were an additional 294 fatal and non-fatal strokes prior to the 1-year landmark.

There were 694 new stroke/TE events in the 3-year follow-up post-landmark date, of which 224 (32.3%) were ischaemic stroke deaths, and 199 (89%) coded as I64. The incidence rate of stroke/TE events by warfarin exposure is shown in [Table 2].

Univariate associations with stroke/TE

Survival to stroke/TE was poorer for those who discontinued warfarin, compared with those not exposed, initiating or persisting, but the difference was not statistically significant (Log Rank 6.17, df 3, p=0.10) [Figure 2]. Women, those with higher CHA2DS2-VASc scores, aged ≥80 years, with a history of bleeding or CKD, and with AF as a secondary diagnosis were at higher risk for stroke/TE.

Multivariate associations with stroke/TE

Figure 2. Kaplan-Meier time to a first fatal or non-fatal stroke/systemic thromboembolism in the 3-year follow-up from the landmark date by warfarin use at landmark

---

---

CHA2DS2-VASc scores ≥5 were independently associated with the risk for stroke/TE, with the adjusted hazard more than doubling [Table 4]. In a sensitivity analysis the associations with the outcome were unchanged by exclusion of 281 events coded as ‘unspecified’ strokes.

There were 2,042 patients (11.8%) with an admission for major bleeding in the 10 years prior to the index admission. Of the 249 patients with a principal diagnosis of bleeding at index admission, the majority were neurological events coded as ICH (n=74, 29.7%), 12 (4.8%) were sub-arachnoid and 33 (13.3%) epidural haemorrhages. A further 88 (35.3%) admissions were for gastro-intestinal bleeding, and 19 (7.6%) had respiratory tract bleeding.

There were 13 deaths (5.2%) within 30 days of admission, and an additional 48 deaths (20.3%) among 236 survivors of a serious bleed within the first year. There were 35 new bleeding events prior to the 1-year landmark.

Of those alive at the landmark date, 608 were hospitalised with a first or new major bleed within three years. The majority were gastrointestinal bleeds (n=410, 67.4%), with 124 ICH 124 (20.4%) and 61 coded as epidural haemorrhages (10%). Of the 22 deaths coded as intracerebral haemorrhages, 21 occurred in hospital. The incidence rate of bleeding events by warfarin exposure is shown in [Table 2].

Univariate associations with major bleeding

The risk for bleeding in the 3-year follow-up was significantly greater for men, those aged ≥80 years, with a prior bleeding event, and with a CHA2DS2-VASc score ≥4. Initiation of and persistence with warfarin were also associated with increased risk (Log Rank 18.01, df 3, p<0.001) [Figure 3].

Figure 3. Kaplan-Meier time to admission for major bleeding within three years of the 1-year landmark date by warfarin use

---

---

Multivariate associations with major bleeding

After adjustment, the independent risk for a serious bleeding event was higher for males, those aged ≥80 years, or with a history of bleeding, and for those both initiating and persisting with warfarin use [Table 5].

This use of linked state-wide morbidity and mortality data and the national pharmaceutical database, allowed a large-scale population-level study with reliable and complete outcome data. Good ascertainment of stroke/TE and serious bleeding events is probable as most cases would require hospital admission. While clinical detail is reduced when using administrative data, it is sufficient for the calculation of the CHA2DS2-VASc score [³⁰], an important predictor of stroke risk [¹⁵].

The externally coded cause of death uses coding conventions which may attribute death to underlying (antecedent or contributing) conditions or to the immediate ‘cause’. The attribution of stroke as cause of death out of hospital to ICD-10-AM I64 (not specified as haemorrhage or infarction) is a worldwide limitation for studies of stroke [³¹,³²,³³].

As the cohort was identified from hospital admissions, the patients may not be representative of the wider population with AF. They are likely to have poorer health and suffer more comorbidity. The proportion of patients with comorbid cardiovascular disease exceeded 40%. We could not quantify the use of anti-platelets (including aspirin) in these patients, a factor in apparent ‘under-use’ of OAC [³⁴]. Use of antiplatelet and OAC drugs among those not in the cohort (not hospitalised for any reason during the study period) may well have differed from those studied.

Exposure to warfarin is difficult to quantify with accuracy from pharmacy data alone, as dose may vary with age, sex and comorbidity [³⁵]. Regimens such as different dosing on alternate days, and varying strength of tablets, make calculation of measures such as a ‘defined daily dose’ impractical. Without INR data, more reliable measures such as ‘time within the therapeutic range’ cannot be assessed [³⁶].

Access to other clinical databases, such as laboratory results, will further strengthen large-scale cohort studies using linked data, with the potential to provide detailed information about the changing use of oral anticoagulants and the resulting outcomes in AF and other thromboembolic conditions.

We thank the following institutions for providing the data used in this study. The Australian Department of Health for the cross-jurisdictional linked PBS data. Staff at the WA Data Linkage Branch and data custodians of the WA Department of Health Inpatient Data Collections and Registrar General for access to and provision of the State linked data. The Victorian Department of Justice and Regulation for the cause of death data held in the National Coronial Information System.

None declared.