The patterns of anticoagulation control and the risk of stroke, bleeding and mortality in patients with non-valvular atrial fibrillation

Authors


Tjeerd P. van Staa, Clinical Practice Research Datalink, Medicines and Healthcare Products Regulatory Agency, 151 Buckingham Palace Road, London SW1W 9SZ, UK.
Tel.: +44 (0) 20 3080 6019 ; fax: +44 (0) 20 3118 9802.
E-mail: tjeerd.vanstaa@cprd.com

Abstract

Summary.  Background: Anticoagulation control is often summarized using the percentage of time spent in a therapeutic range (TTR). This method does not describe the timing and severity of fluctuations in the International Normalised Ratio (INR).Objective: To evaluate whether the TTR method can be improved by considering the patterns of INR over time.Methods: The cohort included adults aged 40+ years with atrial fibrillation (AF) and laboratory records of INR as recorded in the UK Clinical Practice Research Datalink. Statistical clustering techniques based on simple INR measures were used to describe the patterns of INR. Nested case–control studies calculated the odds ratios (ORs) for the risk of stroke, bleeding and mortality with TTR and different INR patterns. It was also evaluated whether cluster analyses improved the prediction of outcomes over TTR.Results: A number of 27 381 patients were studied with a mean age of 73 years. The OR for mortality was 3.76 (95% confidence interval [CI] 3.03–4.68) in patients with < 30% TTR compared with patients with 100% TTR. INR patterns were found to be best described by six different clusters. The cluster with the most unstable pattern was associated with the largest risk of mortality (OR 10.7, 95% CI 8.27–13.85) and stroke (OR 3.42, 95% CI 2.08–5.63). INR measures that predicted death independent of the TTR-included absolute difference between two subsequent INR measurements and change relative to the mean over time.Conclusion: Cluster analysis of INR patterns improved the prediction of clinical outcomes over TTR and may help to identify warfarin users who need additional anticoagulation monitoring.

Introduction

Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia [1]. The condition carries an increased risk of arterial thromboembolism and an ischemic stroke owing to embolization of thrombi that form within the left atrium of the heart [2]. A meta-analysis of randomized controlled trials found that adjusted-dose warfarin reduced the incidence of stroke by 64% (95% confidence interval [CI] 49%–74%) [3,4]. However, an observational study found that the risk of stroke was only reduced by 38% in current compared with past users of warfarin [relative rate (RR) of 0.62, 95% CI 0.54–0.71] [5]. Warfarin use requires frequent blood tests to monitor the level of anticoagulation which is measured by the International Normalized Ratio (INR). Previous reports have found lower risks of stroke at INR levels between 2.0 and 3.5 and an increasing risk with INR values below 2.0 or above 3.5[5–7]. Poor anticoagulation control occurs frequently in AF patients. It has been reported that most AF patients admitted to the hospital with a stroke who were candidates for anticoagulation were either not taking warfarin or had a sub-therapeutic INR at the time of the event [8]. Moreover, several previous studies have found that INR values were out of the target range approximately half the time [9].This poor anticoagulation control can be caused by several reasons including but not limited to non-adherence, drug–drug and drug–food interactions. Furthermore, intercurrent illness or an exacerbation of an existing illness may also contribute to changes in INR levels [10].

Long-term anticoagulation control is often summarized using the percentage of time spent within the therapeutic range (TTR). This method assumes that INR values vary linearly between two measurements [11]. Typically the range recommended by international guidelines is 2.0–3.0 [2]. Several studies have reported that the risk of stroke is increased with a reduced TTR [9,12–15]. A limitation of the TTR method is that the extent that INR values fluctuate over time and the timing and severity of fluctuations are not measured. The objectives of this study were to evaluate whether anticoagulation control can be described by considering the patterns of INR over time and whether these INR patterns improve the prediction of mortality, stroke and bleeding.

Methods

Source of data

This study used data from the General Practice Research Database which is part of the Clinical Practice Research Datalink (CPRD) in the United Kingdom. CPRD comprises the computerized medical records maintained by general practitioners (GPs). GPs play a key role in the UK health care system, as they are responsible for primary health care and specialist referrals. Patients are affiliated with a practice, which centralizes the medical information from the GPs, specialist referrals and hospitalizations. The data recorded in the CPRD since 1987 include demographic information, prescription details, clinical events, preventive care provided, specialist referrals, hospital admissions and their major outcomes [16]. A recent review of validation studies found that medical data in CPRD were generally of high quality [17]. Patients in about 40% of CPRD have now been linked individually and anonymously to the national registry of hospital admission (Hospital Episode Statistics [HES]). For each hospitalized patient, the hospital charts are reviewed, dates of admission and discharge and main diagnoses are extracted, coded by coding staff and collated nationally into HES. The data from HES and CPRD were recorded and collected independently from each other.

Study population

The study cohort included adults aged 40 years or older who had a record of AF and who had their first three INR readings (if recorded in CPRD) within a 6-month period and who were registered with CPRD for at least 6 months prior to the first-ever INR reading (i.e. index date). Cohort assembly took place from 1987 until 2010. Patients were followed from the index date until 6 months after the last INR measured within 6 months of the previous one; if there was a gap of more than 6 months between two subsequent INR measurements, the patient was censored 6 months after the former INR measurement. Patients with a history of heart valve problems and/or replacement surgery were excluded. Patients were included irrespective of the type of AF (paroxysmal or chronic AF).

Longitudinal patterns of INR

INR values of warfarin users typically are recorded in CPRD if the laboratory copies the results electronically to the general practice (as requested by a hospital anticoagulation clinic) or if the practice holds in-house anticoagulation clinics. In order to describe longitudinal anticoagulation patterns, INR readings were ordered by date and those with at least three INR readings in the 6 months after were selected (referred to as 6-months groups). Each 6-months group formed a unit of analysis. A patient could contribute to multiple groups if the INR was measured repeatedly over time. Two approaches were used for describing the longitudinal patterns of INR values of these groups:

  • 1The first approach was based on the percentage of TTR (INR 2.0–3.0) using the Rosendaal method of linear interpolation [11]. This method assumes that INR values vary linearly between two measurements. Using this relationship, the date at which a patient achieved or departed from an INR level within therapeutic range was computed. This was then used to calculate the percentage of their follow-up time spent within the therapeutic range. If there was a measurement after the end of a time period, this was also used in the calculations, in order to obtain an estimate on the time in range from the last INR measurement in the time period up to the end of the time period.
  • 2The second approach of describing anticoagulation patterns was based on two-stage statistical modeling techniques employing the methods described by Leffondréet al. [18]. In the first stage of the modeling, a large number of simple measures were estimated for each of the INR groups including elementary measures of change, measures of non-linearity and of inconsistency of change, measures sensitive to non-monotonicity and to abrupt short-term fluctuations and measures contrasting early versus later change [18]. Examples of these measures include range from minimum to maximum levels, mean-over-time and standard deviation of differences in levels. The percentage of time spent within, below and above the therapeutic range, the mean of INR values below and above normal and the number of INR readings in each 6-month period were also measured. Factor analysis was then used, after Z-score standardization, to eliminate the redundant and correlated measures and those with low variability. Factors with an eigenvalue of > 1.0 were kept. Cluster analysis was then applied to these remaining measures in order to classify each INR group into separate clusters. As the number of clusters needs to be pre-defined in a cluster analysis, we repeated the analyses with this number ranging from 2 to 50. Three criteria were used to determine the most optimal number of clusters, including the observed overall R2 (i.e. proportion of total variance explained by cluster structure), the pseudo-F statistic (i.e. the ratio of the between-clusters mean square to the within-clusters mean square) and the cubic clustering criterion (i.e. comparing the overall R2 to its approximate expected value) [18]. Given the interest in describing patterns with INR readings outside the reference range, INR groups with all values within the normal range were excluded from the factor and cluster analyses.

Outcomes of interest

The following outcomes were analysed: all-cause mortality (based on CPRD), stroke or a transient ischemic attack (TIA) (based on CPRD or HES), major bleed (based on CPRD), minor bleed (based on CPRD) and hospitalization for a bleed (based on HES). A major and minor bleed were defined as agreed by the International Society on Thrombosis and Haemostasis (ISTH) [19]. The analysis based on CPRD date for stroke included all types of stroke (ischemic and hemorrhagic). The reason for this was that most CPRD records of stroke did not contain details of the type of stroke (the GPs tend to record the occurrence of stroke without specifying the etiology). Given the different coding dictionaries used by HES and CPRD and the different methods for data collection, analyses were conducted separately for each source of outcomes. Analyses based on outcomes recorded in HES were restricted to patients from practices participating in the linkage and to those with data during the HES data collection period. Analyses that were based on CPRD outcomes used the complete study population.

Case–control analyses

The effects of INR patterns on the risk of each of the clinical outcomes were evaluated by case–control studies nested within the cohort of AF patients. Two sets of case–control analyses were conducted:

  • 1In the first set of analyses, each case of the selected outcome was matched to controls by practice, gender, age (within 5 years), calendar year (within 2 years) and duration of time since first-ever INR reading. If no controls were found, controls from other practices were selected. Both cases and controls were required to have a 6-month INR group that had at least one INR measurement in the 3 months before (as summarized in Fig. 1). The set of analyses was not matched by TTR in order to evaluate the effects of TTR and the pattern of the most recent INR readings on the risk of the clinical outcomes.
  • 2In the second set of analyses, cases and controls were matched additionally by percentage of TTR of the most recent 6-month INR group (within a maximum of 10%). This set of analyses evaluated whether any of the simple measures as identified in the cluster analyses contributed to the risk of clinical outcomes independently of TTR.
Figure 1.

 Methods for classifying International Normalized Ratio (INR) measurements in cases and controls.

For both case–controls analyses, the INR groups with 100% TTR constituted the reference group.

The number of controls varied by outcome, balancing close matching and ability to find sufficient numbers of controls: one control was selected for each case who died, three controls for each case with major and minor bleeds and six controls for all other outcomes. Conditional logistic regression was used to estimate the odds ratios (ORs). The ORs were adjusted for the following risk factors [13]; smoking history, body mass index (BMI), socioeconomic status of practice, alcohol use, prescribing in the 6 months before the index date of antiarrhythmics, ACE inhibitors, aspirin, oral corticosteroids and beta-blockers and a history prior to the index date of bleeds, cancer, cerebrovascular disease, congestive heart failure, ischemic heart disease, chronic obstructive pulmonary disease, diabetes mellitus, hypertension, liver and renal impairment, peripheral vascular disease and venous thromboembolism. Missing data on smoking history, BMI and/or alcohol use were handled by including an indicator of missingness into the regression analyses.

Results

The study population included 27 381 patients with a median duration of follow-up of 3 years and a mean age of 73 years. Baseline characteristics of the overall study population are described in Table 1. Table 2 shows the characteristics of the cases and controls that were selected from the overall study population. The majority of cases who died were matched to controls by year of birth, gender and practice.

Table 1.   Baseline characteristics of the study population
 Total (n = 27 381)
Follow-up (years)
 Mean (standard deviation)4 (3)
 Median3
Age (years)
 Mean (standard deviation)73 (10)
 Median75
Male (%)15 890 (58.0)
Medical history (%)
 Diabetes mellitus4430 (16.2)
 Hypertension20 133 (73.5)
 Bleeds6443 (23.5)
 Cancer3150 (11.5)
 Cerebrovascular disease5619 (20.5)
 Congestive heart failure4672 (17.1)
Table 2.   Baseline characteristics of cases (of death, stroke and major bleed) and controls
 Death*Stroke*Major bleed*
CasesControlsCasesControlsCasesControls
  1. *As reported in Clinical Practice Research Datalink (CPRD).

Age (years)
 Mean (standard deviation)80.1 (8)79.8 (7.8)78.3 (8.7)78.2 (8.5)75.9 (8.4)75.9 (8.3)
 Median818179797777
Male (%)2365 (58.7)2365 (58.7)273 (58.3)1638 (58.4)964 (59.5)2892 (59.5)
Medical history (%)
 Diabetes mellitus750 (18.6)524 (13)77 (16.5)358 (12.8)236 (14.6)712 (14.6)
 Bleeds971 (24.1)882 (21.9)107 (22.9)597 (21.3)231 (14.3)545 (11.2)
 Cancer594 (14.7)453 (11.2)49 (10.5)296 (10.6)181 (11.2)468 (9.6)
 Cerebrovascular disease1052 (26.1)959 (23.8)89 (19)335 (12)327 (20.2)1020 (21)
 Congestive heart failure1328 (33)714 (17.7)83 (17.7)513 (18.3)276 (17)841 (17.3)
 Hypertension2981 (74)2837 (70.4)333 (71.2)2010 (71.7)1151 (71)3467 (71.3)

Table 3 shows the association between the risk of stroke, major/minor bleed and death with levels of TTR. We found a strong association between TTR and risk of mortality, stroke and bleeding. Patients who spent < 30% in TTR had a higher risk of developing a stroke (OR 2.6, 95% CI 1.7–4.0) compared with those patients with 100% TTR. A higher risk was also found for overall mortality which showed an OR of 3.8 (95% CI 3.0–4.7) in patients with < 30% TTR.

Table 3.   Odds ratio (OR) for clinical outcomes for different levels of the percentage of time in therapeutic range (TTR)
OutcomeTTR (%)Number of casesNumber of controlsCrude OR (95% CI)Adjusted OR (95% CI)*
  1. *OR adjusted for smoking history, body mass index, socio-economic status of practice, alcohol use, prescribing of medication 6 months before the index date and history of co-morbidities prior to the index date.

  2. CPRD, Clinical Practice Research Datalink; HES, Hospital Episode Statistics.

As reported in CPRD< 306494304.3 (3.5–5.2)3.8 (3.0–4.7)
Death30–394561906.0 (4.8–7.5)5.5 (4.3–7.0)
40–495203344.0 (3.3–4.9)3.5 (2.8–4.4)
50–596124283.7 (3.0–4.4)3.4 (2.7–4.1)
60–695645642.6 (2.1–3.1)2.4 (2.0–3.0)
70–794315602.0 (1.6–2.4)1.8 (1.5–2.3)
80–893304791.7 (1.4–2.1)1.7 (1.4–2.1)
90–991823391.4 (1.1–1.8)1.4 (1.1–1.8)
100284704ReferenceReference
Stroke< 30602332.7 (1.7–4.0)2.6 (1.7–4.0)
30–39401552.5 (1.6–4.0)2.4 (1.5–3.9)
40–49402371.6 (1.0–2.5)1.6 (1.0–2.5)
50–59603241.8 (1.2–2.6)1.8 (1.2–2.7)
60–69593781.5 (1.0–2.3)1.3 (0.9–2.0)
70–79673401.9 (1.3–2.8)1.9 (1.3–2.8)
80–89533841.3 (0.9–2.0)1.3 (0.9–2.0)
90–99382641.4 (0.9–2.1)1.4 (0.9–2.1)
10051488ReferenceReference
Minor bleed< 401473451.6 (1.2–1.9)1.6 (1.2–2.0)
40–592025671.3 (1.0–1.6)1.8 (1.3–2.4)
60–792878111.3 (1.1–1.5)1.3 (1.1–1.5)
≥ 803461220ReferenceReference
Major bleed< 402265821.3 (1.0–1.5)1.4 (1.0–1.5)
40–593089151.1 (0.9–1.3)1.1 (0.9–1.3)
60–7944713401.1 (0.9–1.2)1.1 (0.9–1.2)
≥ 806402026ReferenceReference
As reported in HES< 40371751.4 (0.9–2.1)1.3 (0.8–2.0)
Bleed40–59492361.3 (0.9–1.9)1.3 (0.9–2.0)
60–79443540.8 (0.5–1.1)0.8 (0.5–1.2)
≥ 8079489ReferenceReference
Stroke< 40441652.5 (1.5–4.0)2.4 (1.5–4.0)
40–59492171.9 (1.3–3.0)2.1 (1.3–3.2)
60–79403001.1 (0.7–1.8)1.2 (0.8–1.9)
≥ 8048404ReferenceReference

From the factor analysis, eight measures were kept and used in the cluster analysis. The following basic descriptive measures of change were used: percentage spent below and above the therapeutic range, the mean of INR values above normal (> 3.0) and the number of INR readings in each 6-month period. Some elementary measures of change were also included, such as change between the first and the last measurement and the slope of the simple linear model. These measures provide information about the evolution between the first and the last INR measurement, for example an increase or decrease over time. The maximum of the absolute differences between two subsequent measurements was also included, which reveals the presence of an important increase or decrease between two subsequent INR measurements. The cluster analysis found that a total of 10 different clusters best described the INR patterns over time. Out of these 10 clusters, five occurred infrequently and were combined into the ‘rest cluster’ (cluster 6). Visual review of INR patterns in the ‘rest cluster’ showed that INR values were heavily fluctuating and had no consistency.

Table 4 describes the frequencies of the six INR clusters and the INR measures that were used to classify into the clusters. The most frequently occurring cluster was cluster 2 (25.2%). Most unstable patterns were found in cluster 5 (17.0%) and cluster 6 (0.2%). A prominent feature of these clusters was the high number of INR measurements in the 6-month period and the positive change between the first and the last measurement relative to the mean. Cluster 6 included INR groups with higher percentages spent above therapeutic range and with a high mean of INR values above therapeutic range. A high maximum of the absolute differences between two consecutive measurements indicates that there is at least one abrupt change present in the pattern. Thus, cluster 6 can be described as extremely fluctuating and with INR values mainly above therapeutic range. Clusters 3 and 4 both showed a negative change between the first and the last INR measurement relative to the mean. This means that patterns assigned to these clusters showed a decrease of INR values over time. These two clusters were discriminated by their value for percentage of INR measurements below therapeutic range, which is lowest in cluster 3 and highest in cluster 4. The INR measures in clusters 1 and 2 indicated stable patterns.

Table 4.   Characteristics of International Normalized Ratio (INR) measures in the six clusters
MeasureReference cluster (N = 145874 [13.8%])Cluster 1 (N = 198003 [18.8%])Cluster 2 (N = 265821 [25.2%])Cluster 3 (N = 173865 [16.5%])Cluster 4 (N = 90444 [8.6%])Cluster 5 (N = 178724 [17.0%])Cluster 6 (N = 1755 [0.2%])
Mean (SD)Mean (SD)Mean (SD)Mean (SD)Mean (SD)Mean (SD)Mean (SD)
  1. *Max |Δ1| = Maximum of the absolute difference between two subsequent INR measurements. Ratio mean |Δ2|/mean |Δ1| = ratio of the mean absolute second differences to the mean of absolute first differences.

Max |Δ1|*0.47 (0.21)0.91 (0.40)1.06 (0.48)1.64 (0.71)2.18 (1.00)1.88 (0.76)5.13 (2.17)
Change relative to mean over time0 (0.15)0.07 (0.34)0.04 (0.26)−0.012 (0.29)−0.71 (0.38)0.07 (0.34)0.12 (0.71)
Percentage below therapeutic range0 (0)0.16 (0.17)0.26 (0.23)0.04 (0.07)0.30 (0.20)0.26 (0.16)0.20 (0.18)
Percentage above therapeutic range0 (0)0.09 (0.11)0.04 (0.06)0.40 (0.19)0.15 (0.12)0.16 (0.11)0.43 (0.26)
Mean of INR values above therapeutic range0 (0)0.05 (0.07)0.03 (0.05)0.23 (0.17)0.20 (0.18)0.12 (0.10)0.98 (0.89)
Slope b of linear line0 (0)0 (0.01)0 (0.01)0 (0.01)−0.01 (0.02)0 (0.01)−0.01 (0.15)
Number of INR measurements4.64 (1.66)7.26 (2.39)7.46 (2.40)8.00 (2.84)11.26 (3.89)14.94 (3.78)12.50 (5.90)
Ratio mean |Δ2|/mean |Δ1|1.69 (0.49)1.27 (0.28)1.87 (0.21)1.78 (0.27)1.53 (0.25)1.63 (0.19)1.66 (0.31)

Figure 2 shows a graphical representation of each cluster for randomly selected patients accompanied with the ORs for death, stroke and major and minor bleeds for each cluster. The most unstable cluster (cluster 6) showed a 10-fold increased risk of mortality (OR 10.7, 95% CI 8.27–13.85) and a substantially increased risk of a stroke (OR 3.42 95% CI 2.08–5.63), major bleed (OR 1.60, 95% CI 1.13–2.26) and minor bleed (OR 2.13, 95% CI 1.39–3.27). ORs were generally lower in clusters with more stable INR patterns over time.

Figure 2.

 Clusters with distinct International Normalized Ratio (INR) patterns and adjusted odds ratios (ORs) with each cluster for death, stroke, major and minor bleed.

The results of the case–control analysis with matching on TTR are shown in Table 5. Differences between two subsequent INR measurements predicted the risks of mortality and stroke independent of TTR. Other INR measures that predicted death independent of TTR were the mean of INR values above therapeutic range, change between the first and the last measurement, the number of INR measurements and percentage above the therapeutic range.

Table 5.   International Normalized Ratio (INR) measures and prediction of death, stroke and bleeding independent of time in therapeutic range (TTR) (i.e. cases and controls were matched on TTR)
OutcomeINR measureOR (95% CI)
DeathMaximum of the absolute difference between two subsequent INR measurements1.60 (1.46–1.76)
Mean of INR values above therapeutic range1.18 (1.07–1.31)
Change relative to the mean over time1.14 (1.08–1.20)
Number of INR measurements1.12 (1.05–1.19)
Percentage above therapeutic range1.08 (1.01–1.16)
StrokeMaximum of the absolute difference between two subsequent INR measurements1.20 (1.09–1.32)
Major bleedMean of INR values above therapeutic range1.12 (1.04–1.20)
Change relative to the mean over time1.09 (1.03–1.15)
Minor bleedMean of INR values above therapeutic range1.17 (1.07–1.28)
Change relative to the mean over time1.16 (1.07–1.25)

A sensitivity analysis was performed to determine the influence of incomplete matching of cases and controls on practice. Similar results were observed for death, minor and major bleeds (as reported in CPRD) and hospitalization for a bleed (as reported in HES). In patients with TTR < 30%, the OR for stroke (as reported in CPRD) was somewhat different in the analysis of subjects with complete matching on practice compared with the main analysis (OR 3.43 95% CI 1.82–6.46 vs. OR 2.6, 95% CI 1.7–4.0). However, this difference was not statistically significant.

Discussion

This study found that longitudinal anticoagulation measurements can be classified into distinct clusters that are associated with the risk of stroke, minor/major bleed and mortality. Patients with unstable anticoagulation patterns had increased risks of stroke, major/minor bleeds and death compared with patients with a stable pattern and good anticoagulation. We confirmed that TTR, as proposed by Rosendaal et al., describes well the risk of major clinical outcomes. Nonetheless, we found that the prediction of death and stroke can also be improved by considering specific measures of change in INR values.

The majority of studies on the quality of anticoagulation control and risk of adverse outcomes, such as stroke and bleeding, have used TTR. The results in the present study found results consistent with these previous studies [9,13–15,20–26]. Although TTR is an uncomplicated measure that is easily interpretable, it does come with some limitations. First, it has a limited ability to describe the extent that values are out of range. Values that are ‘very’ out of range are not distinguished from values that are ‘little’ out of range. Second, the TTR is at best an approximation for the real pattern of the INR as it assumes a linear relation between two INR measurements which may not be correct. There may be a large variability in INR. This was demonstrated in a study of Hylek et al. in which cases had INR values of at least 6 at the time of the stroke but the INR had been within the therapeutic range only a few weeks before [27]. A small number of other studies have proposed other methods for measuring INR control [28,29]. Lind et al. proposed using the standard deviation of transformed INR (SDTINR) for the prediction of clinical outcomes [30]. We think it is of importance to explore other methods other than TTR to describe INR variability in order to improve the prediction of mortality, stroke and bleeding in patients treated with warfarin.

Warfarin has shown to be a very effective medicine in preventing an ischemic stroke in patients with atrial fibrillation. However, anticoagulation control is an important factor that influences the effectiveness and safety of warfarin treatment. Therefore it is of importance, when assessing the benefit-risk balance of warfarin, to take the quality of anticoagulation control into account. In day-to-day anticoagulation control, it may be very useful for a physician to quantify specific changes in INR patterns in order to identify patients who are not well controlled with warfarin and will have a higher risk on getting a bleed or an ischemic stroke. Several INR measures were found in the present study to improve the prediction of major clinical outcomes. However, it must be noted that the measures of change that show a significant better prediction than TTR are specific for this set of data and cannot be extrapolated to other settings without validation. Other measurements of change may be more relevant in different populations. The predictive value of these INR measures will need to be validated in independent studies.

There are several limitations and strengths to this study. This is an observational study and patients were not randomized to each of the different INR patterns. It could also not be established whether an unstable INR pattern caused an increased risk of mortality and stroke or whether instability in INR measurements was a symptom of intercurrent illness. But these findings highlight the importance of increased attention to these patients. We also did not have information on the diagnostic criteria used for the AF diagnosis, such as electrocardiograms. Nonetheless, a previous CPRD study reported a high level of validity in the recording of AF by GPs [1].We were also unable to classify most stroke cases for the underlying etiology (ischemia or bleeding) as most strokes were recorded in CPRD without these details. The procedures used to diagnose strokes (such as CT or MRI) are also not routinely recorded in CPRD and there was no information on the clinical presentation in the stroke cases. A further limitation was that we are not able to verify that the measurement was accurate and that interlaboratory precision is guaranteed. Guidelines do exist to ensure consistency of INR measurement between laboratories in the UK. We did a sensitivity analysis by removing cases and controls not matched by practice to determine the influence of using a different laboratory for INR measurement. The results of this analysis did not show important differences with the analysis including all cases and controls. Lastly, another limitation was that we did not have information on INR values for all patients as not all anticoagulation clinics report these electronically to the GPs. Strengths of the present study are the availability of a large amount of high-quality data, including data on important risk factors. Additionally, we used multiple data sources for stroke and bleeding events, validating these outcomes.

In conclusion, unstable INR patterns are associated with an increased risk of mortality, stroke and bleeding. We assessed that the Roosendaal method for measuring long-term anticoagulation control can be improved by also measuring the magnitude and timing of deviations of INR values from the reference range. Cluster analysis of INR patterns improved prediction of clinical outcomes over TTR and may help to identify warfarin users who need additional anticoagulation monitoring.

Acknowledgements

The authors would like to thank Professor Dr D. E. Singer for his valuable comments on the manuscript. The research leading to these results was conducted as part of the PROTECT consortium (Pharmacoepidemiological Research on Outcomes of Therapeutics by a European ConsorTium, http://www.imi-protect.eu) which is a public–private partnership coordinated by the European Medicines Agency. The views expressed in this paper are those of the authors and do not reflect the official policy or position of the Medicines and Healthcare Products Regulatory Agency, UK.

Disclaimer

The processes described and conclusions drawn from the work presented herein relate solely to the testing of methodologies and representations for the evaluation of benefit and risk of medicines. This report neither replaces nor is intended to replace or comment on any regulatory decisions made by national regulatory agencies, nor the European Medicines Agency.

Funding

The PROTECT project has received support from the Innovative Medicine Initiative Joint Undertaking (http://www.imi.europa.eu) under Grant Agreement n° 115004, resources of which are composed of financial contribution from the European Union’s Seventh Framework Programme (FP7/2007–2013) and EFPIA companies in kind contribution. CPRD is owned by the Secretary of State of the UK Department of Health and operates within the Medicines and Healthcare Products Regulatory Agency (MHRA). CPRD has received funding from the MHRA, Wellcome Trust, Medical Research Council, NIHR Health Technology Assessment Programme, Innovative Medicine Initiative, UK Department of Health, Technology Strategy Board, Seventh Framework Programme EU, various universities, contract research organisations and pharmaceutical companies. The Department of Pharmacoepidemiology and Clinical Pharmacology, Utrecht Institute for Pharmaceutical Sciences, has received unrestricted research funding from the Netherlands Organisation for Health Research and Development (ZonMW), the Dutch Health Care Insurance Board (CVZ), the Royal Dutch Pharmacists Association (KNMP), the private–public funded Top Institute Pharma (http://www.tipharma.nl, includes co-funding from universities, government, and industry), the EU Innovative Medicines Initiative (IMI), EU 7th Framework Program (FP7), the Dutch Medicines Evaluation Board, the Dutch Ministry of Health and industry (including GlaxoSmithKline, Pfizer, and others).

Disclosure of Conflict of Interest

The authors state that they have no conflict of interests.

Ancillary