Best quality indicator of vitamin K antagonist therapy to predict mortality and bleeding in haemodialysis patients with atrial fibrillation

Abstract

Background

There is a high prevalence of atrial fibrillation (AF) in patients undergoing haemodialysis. Oral anticoagulant therapy with vitamin K antagonists (VKAs) is the only accepted treatment for the prevention of thromboembolism in haemodialysis patients with AF. However, in this population, the risk of bleeding is greatly increased. The aim of the study was to evaluate the ability of treatment quality indicators of VKA therapy to predict mortality and bleedings in a population of haemodialysis patients with AF.

Materials and methods

A total of 129 patients were included in this cohort study. Deaths and bleeding events were recorded during a follow-up of 4 years. In all patients, International Normalized Ratio (INR) values were assessed at least once a month. Time in therapeutic range (TTR) and INR variability, as measured by the standard deviation of INR, were updated at each INR measurement. A Cox model with time-dependent co-variates and sandwich variance was applied.

Results

During follow-up, 71 patients died and 55 bleeding episodes occurred in 31 patients. INR variability was the only indicator associated with both mortality (hazard ratio [HR]=1.67, 95% confidence interval [CI] 1.12; 2.49, p=0.012) and bleeding (HR=2.85, 95% CI: 1.71; 4.75, p=0.0001). HR of mortality was higher in patients with INR >3 (HR=2.06, 95% CI: 1.09; 3.88, p=0.0259) than in subjects in therapeutic range 2<INR≤3. TTR was inversely associated with the risk of recurrent haemorrhagic events (HR=0.88, 95% CI: 0.80; 0.95, p=0.0023), but not with a first episode of bleeding. Results were consistent after censoring patients at VKA withdrawal.

Discussion

Our study suggests that, in haemodialysis patients with AF taking VKAs, INR variability is the quality indicator that best predicts clinical outcomes. In this population, if more treatment quality indicators are considered together, it may become easier to identify patients at particularly high risk of bleeding and death.

INTRODUCTION

Atrial fibrillation (AF) is very frequent in patients with end-stage renal disease and its high prevalence is associated with an increased incidence of stroke¹. Oral anticoagulant treatment is the cornerstone of stroke prevention in AF patients; the most recent cardiology guidelines suggest direct oral anticoagulants (DOACs) as the first choice and vitamin K antagonists (VKAs) as an alternative in some patients. The same guidelines discourage the use of DOACs in end-stage renal disease patients undergoing haemodialysis (HD) and suggest the use of VKAs as the only available alternative². However, these particular patients have a higher risk of bleeding due to their intrinsic frailty and alterations in platelet function associated with uraemia³. Moreover, in patients on VKAs, an inverse correlation has been shown between renal function and the quality of the anticoagulant treatment as expressed by the time in therapeutic range (TTR)⁴. This means that HD patients spend less time within the planned International Normalized Ratio (INR) intervals, and have a higher risk of over-anticoagulation and of major bleedings compared to patients with preserved kidney function or with mild to moderate renal dysfunction^5,6. It has also been shown that, in HD patients, treatment with VKAs results in a 4-fold increase in the risk of bleeding⁷. For these reasons, it is extremely problematic for nephrologists to manage oral anticoagulants in the end-stage renal disease (ESRD) population, and only a minority of HD patients with AF receive anticoagulant treatment even if this would be strongly recommended by most guidelines on the basis of the Atrial Fibrillation Stroke Risk score (CHA₂DS₂) and the revised CHA₂DS₂VASc score^8,9. Few data about the actual quality of the anticoagulant treatment (i.e., studies reporting the INR and TTR values) in HD patients are available. The aim of our study is to evaluate which treatment quality indicators of VKA therapy could be useful in predicting mortality and haemorrhagic events in a population of patients with AF undergoing HD and taking VKAs. In particular, we evaluated the association of the longitudinal values of INR, TTR and INR variability (as measured with standard deviation of INR, SD[INR]) with mortality and bleedings in this setting.

MATERIALS AND METHODS

Study design

All patients under observation in ten Italian Dialysis Centres on 31/10/2010 were considered (n=1,529) and their clinical charts were reviewed for eligibility for the study. Two-hundred and ninety patients with at least one documented paroxysmal or persistent AF episode, or with permanent AF were recruited. Different types of AF were defined in agreement with the European Society of Cardiology guidelines¹⁰. Only patients treated with VKAs with at least two INR measures (n=129) were included in the study (Figure 1). At recruitment (31/10/2010), data were collected on the presence of hypertension (systolic blood pressure ≥140 mmHg and/or diastolic blood pressure ≥90 mmHg before the beginning of the HD session or use of anti-hypertensive drugs), diabetes mellitus, ischaemic heart disease (previous myocardial infarction or coronary revascularisation procedures and/or previous hospitalisation due to acute coronary syndrome), heart failure (presence of left-ventricular ejection fraction <50% and/or previous hospitalisation due to acute or chronic heart failure), previous strokes (defined as ischaemic or haemorrhagic by computed tomography scan or magnetic resonance imaging), and major bleeding episodes (haemorrhagic episodes requiring hospitalisation or blood transfusion, or causing a >2 g/dL reduction in haemoglobin, or involvement of a critical anatomical site (intracranial, spinal, ocular, pericardial, articular, intramuscular with compartment syndrome, retroperitoneal¹¹) and on use of antiplatelets (aspirin, ticlopidine and clopidogrel).

Figure 1. Flow-chart of patients included in the study.

INR: International Normalized Range.

Patients were prospectively followed up for 4 years (from 31/10/2010 until 31/10/2014 or death) and their medical records were updated at each dialysis session. Deaths from any cause and major bleeding events occurring during the follow-up were recorded. In all patients, the INR values were assessed at least once a month. In order to assess the quality of treatment in patients on VKAs, we selected and calculated the following three indices: INR value, TTR (according to Rosendaal)¹², and the standard deviation of INR, SD(INR)¹³. Thromboembolic and haemorrhagic risks were calculated using the CHA₂DS₂VASc score and the HAS-BLED score for Major Bleeding Risk¹⁰.

Procedures were performed according to the Declaration of Helsinki and approved by the local ethics committee. Informed consent was obtained before enrolment on the study.

Statistical analysis

Categorical data were described by counts and percentages, and compared by the χ² test, while continuous variables were described by medians and quartiles, and compared by the Mann-Whitney U-test. Incidence rates of mortality and bleeding were computed as the total number of events divided by the total patient-years at risk. The Poisson model was used to model mortality and bleeding rates by INR values, TTR and SD(INR), by the use of polynomials, the degree of which was chosen by the likelihood ratio test. The survival time was computed as the difference between the date of death and the start of the study (recruitment). Subjects were censored at their last follow-up visit or at the end of the study. The Kaplan-Meier estimator was used to estimate survival of the patients during follow-up.

To account for the longitudinal nature of the markers (INR, TTR and SD[INR]), we applied the Cox regression model with time-dependent co-variates to estimate the association of these longitudinal markers with mortality and bleeding (both first and recurrent bleedings). A robust variance estimate (sandwich estimator) was used to account for within-subject correlation, with particular consideration of recurrent episodes of bleeding, with a generalised estimating equation (GEE) approach. Proportional hazard assumption was checked by Schoenfeld residuals, and variables not fulfilling this were used as strata (i.e., age in models on bleeding outcome) in a stratified Cox model. Given the U-shaped relationship between INR and mortality, INR was modelled with a third-degree polynomial and was also treated as a categorical variable, with values below, within and above the recommended therapeutic range (<2, 2–3, >3). When dealing with the outcome of bleeding, INR was included as a continuous variable with a linear effect. TTR was modelled both as a continuous variable with a linear effect and categorised as higher and lower than 65%, while SD (INR) was treated only as a continuous variable. Co-variates included in the models were selected from a penalised Lasso model considering all baseline variables reported in Table I. Shrinkage parameter (penalisation) was chosen to maximise the cross-validated log-likelihood¹⁴. For mortality, the models were adjusted for: dialytic age, heart failure, permanent AF, and HAS-BLED and CHA₂DS₂VASc scores. For bleeding, the models were stratified by age (< or ≥75 years) and adjusted for dialytic age and heart failure. We performed both an “intention to treat” analysis considering the whole follow-up of all patients and an “as treated” analysis censoring subjects at their VKA withdrawal. Early events occurring within 30 days of VKA withdrawal were also included in the “as treated” analysis.

Table I.

Baseline characteristics of patients included in the study and unadjusted association with mortality and bleeding events

Characteristics	Overall (n=129)	Alive (n=58)	Dead (n=71)	Test	No bleeding (n=98)	At least 1 bleeding (n=31)	Test
	N (%)	N (%)	N (%)	p	N (%)	N (%)	p
Gender: Female	44 (34)	19 (33)	25 (35)	0.916	33 (34)	11 (35)	1
Age (years) median [IQR]	75.2 [68.4, 79.9]	71.9 [63.9, 79.8]	76.6 [71.6, 80.5]	0.017	76.5 [69.7, 79.9]	71.8 [66.5, 79.1]	0.084
Age ≥75 (years)	65 (50)	25 (43)	40 (56)	0.187	54 (55)	11 (35)	0.089
Dialytic age median [IQR]	3.6 [1.5, 7.2]	5.5 [3.0, 9.3]	2.9 [1.2, 6.4]	0.004	3.2 [1.4, 6.1]	8.3 [4.8, 11.5]	<0.001
Dialytic age ≥3 (years)	77 (60)	43 (74)	34 (48)	0.004	52 (53)	25 (81)	0.012
Hypertension	97 (75)	43 (74)	54 (76)	0.963	72 (73)	25 (81)	0.57
Diabetes mellitus	37 (29)	14 (24)	23 (32)	0.403	30 (31)	7 (23)	0.526
Ischaemic heart disease	57 (44)	23 (40)	34 (48)	0.448	45 (46)	12 (39)	0.619
Heart failure	55 (43)	16 (28)	39 (55)	0.003	46 (47)	9 (29)	0.121
Ischaemic stroke	19 (15)	6 (10)	13 (18)	0.308	15 (15)	4 (13)	0.969
Bleeding/Haemorrhagic stroke	16 (12)	7 (12)	9 (13)	1	12 (12)	4 (13)	1
Permanent atrial fibrillation	66 (51)	23 (40)	43 (61)	0.029	49 (50)	17 (55)	0.792
Antiplatelet therapy	24 (19)	8 (14)	16 (23)	0.297	19 (19)	5 (16)	0.887
CHA 2 DS 2 VASc score median [IQR]	4.0 [3.0, 5.0]	4.0 [3.0, 5.0]	4.0 [4.0, 5.5]	0.002	4.0 [3.0, 5.0]	4.0 [3.0, 4.0]	0.05
CHA 2 DS 2 VASc score
0–1	3 (2)	3 (5)	0 (0)		2 (2)	1 (3)
2–9	126 (98)	55 (95)	71 (100)		96 (98)	30 (97)
HAS-BLED score median [IQR]	3.0 [3.0, 4.0]	3.0 [2.0, 3.0]	3.0 [3.0, 4.0]	0.012	3.0 [3.0, 4.0]	3.0 [2.5, 3.0]	0.623
HAS-BLED score
0–2	32 (25)	19 (33)	13 (18)		24 (24)	8 (26)
3–9	97 (75)	39 (67)	58 (82)		74 (76)	23 (74)
INR at recruitment median [IQR]	2.2 [1.7, 2.5]	2.2 [1.7, 2.6]	2.2 [1.8, 2.5]	0.953	2.2 [1.8, 2.6]	2.0 [1.7, 2.5]	0.217

N: number; y: years; IQR: interquartile range; INR: International Randomized Range.

Results of the Cox models are expressed in terms of estimated hazard ratios (HR), 95% confidence intervals (95% CI) and p-values. Analyses were carried out using R statistical software version 3.4.4 (2018-03-15) (The R Foundation).

RESULTS

Table I shows the clinical characteristics of the study population (n=129) at recruitment. The median age at recruitment was 75 years (1^st and 3^rd quartiles: 68, 80) and 85 (66%) were males. Most of the patients had a long history of renal replacement therapy and numerous comorbidities. Nineteen (15%) patients had a previous experience of a thromboembolic event and 16 (12%) had suffered a major bleeding episode. Half of the patients showed persistent AF. Both thromboembolic (CHA₂DS₂VASc) and haemorrhagic (HAS-BLED) scores were high. Almost all patients had a CHA₂DS₂VASc score >2, and therefore had a strong indication for oral anticoagulant therapy, and 98 (75%) had HAS-BLED scores ≥3, and could thus be considered at high risk of bleeding episodes. Median INR at recruitment was 2.2 (1^st and 3^rd quartiles: 1.8, 2.5).

Median follow-up was 48 months (1^st and 3^rd quartile: 48 months, range 2–48 months). A total of 39 patients interrupted VKA therapy before the end of the study. A total of 9,730 INR measurements were available during the follow-up of the 129 patients, with a median of 68 measurements per subject (range 2–222). Median INR was 2.2 (1^st and 3^rd quartiles: 1.8, 2.7). Overall median TTR was 55.5% (1^st and 3^rd quartiles: 40.9%, 67.2%). Overall median SD(INR) was 0.61 (1^st and 3^rd quartiles: 0.47, 0.86).

Mortality

During follow-up, 71 patients died (survival estimate at 4 years 0.43, 95% CI: 0.35, 0.52). Age, dialytic age, heart failure, permanent atrial fibrillation, and the CHA₂DS₂VASc and HAS-BLED scores were all significantly associated with mortality (Table I). All causes of mortality are shown in Table II.

Table II.

Causes of deaths (n=71)

Cause of death	N (%)
Sudden cardiac death	9 (13%)
Cardiogenic shock	6 (8%)
Sepsis	14 (20%)
Neoplasia	6 (8%)
Cachexia	21 (30%)
Ischaemic stroke	2 (3%)
Haemorrhagic stroke	2 (3%)
Other	11 (15%)

The absolute rate of mortality as a function of INR, TTR and SD(INR) is shown in Figure 2. Mortality rate showed a non-linear U-shaped relationship with INR values (Figure 2A), while it decreased with higher values of TTR (Figure 2B) and increased with increasing INR variability (Figure 2C).

Figure 2. Mortality and bleeding rates (per person-years, thick lines) vs longitudinal values of International Normalized Range (INR), time in therapeutic range (TTR) and standard deviation INR (SD(INR)) with 95% confidence intervals (shaded area).

Mortality rate vs INR (A), TTR (B) and SD (INR) (C). Bleeding rate vs INR (D), TTR (E) and SD(INR) (F).

The analysis of the association among the longitudinal indicators and mortality, adjusted for relevant covariates, is reported in Table III. This showed a higher mortality for patients with INR >3 as compared with those with INR values between 2 and 3 (HR=2.06, 95% CI: 1.09; 3.88, p=0.0259), but this difference lost statistical significance in the “as treated” analysis (Online Supplementary Table SI). When INR was modeled as a continuous variable with a third-degree polynomial, it resembled the typical U-shaped form (data not shown) and its overall impact reached statistical significance only in the “as treated” analysis (p=0.0361) (data not shown). TTR was not associated with mortality, while the INR variability showed a significant association, increasing the hazard of mortality by 67% for each unit increment of SD(INR) (HR=1.67, 95% CI: 1.12; 2.49, p=0.0337). This association remained and appeared stronger when patients were censored at VKA withdrawal (Online Supplementary Table SI).

Table III.

Results of the Cox model analysis of the association between International Normalized Range (INR), time in therapeutic range (TTR), and INR variability and mortality for any cause, adjusted by relevant co-variates (n=129, 71 deaths)

Model*	Predictors	Intention to treat analysis§ 71 events
		Hazard Ratio (95% CI)	p
(a2)	INR<2	1.14 (0.69; 1.89)	0.5996
	2≤INR≤3	Ref
	INR>3	2.06 (1.09; 3.88)	0.0259
(b)	TTR, per 10% increase	0.95 (0.85; 1.07)	0.3954
(b2)	TTR ≥65% vs <65%	0.77 (0.42; 1.4)	0.3863
(c)	SD (INR), per 1 unit increase	1.67 (1.12; 2.49)	0.012

^*Adjusted for: dialytic age, heart failure, permanent atrial fibrillation (AF), CHA₂DS₂VASc and HAS-BLED.

^§Co-variates statistically significant in intention to treat analysis: (a) dialytic age (p=0.023), heart failure (p=0.0293), and HAS-BLED (p=0.0174); (b) dialytic age (p=0.0397), heart failure (p=0.0183), and HAS-BLED (p=0.0129); (c) dialytic age (p=0.0307), heart failure (p=0.0229), and HAS-BLED (p=0.0163); (d) dialytic age (p=0.0405), heart failure (p-value=0.0132), and HAS-BLED (p=0.0182).

Bleeding episodes

During follow-up, 55 bleeding episodes occurred in 31 patients with an overall rate per person-year of 0.170 (95% CI: 0.128, 0.222). Only dialytic age and the CHA₂DS₂VASc score were associated with the occurrence of at least one bleeding episode (Table I).

The absolute rate of recurrent bleedings as a function of INR, TTR and SD(INR) is shown in Figure 2. The rate of recurrent bleeding episodes increased as a function of INR (Figure 2D) and its variability (SD[INR]) (Figure 2), and decreased as a function of TTR (Figure 2E).

Table IV shows the association between the three longitudinal indicators and haemorrhagic events. There was no statistical difference in the association between INR and the first or recurrent bleedings when INR was treated as categorical or as continuous variable.

Table IV.

Results of the Cox model analysis of the association between International Normalized Range (INR), time in therapeutic range (TTR), and INR variability (SD(INR)), and both first and recurrent bleedings, adjusted by relevant co-variates (n=129, 31 first bleedings and 55 recurrent bleedings)

Model*	Predictors	First bleeding Intention to treat analysis§ 31 events
		Hazard ratio (95% CI)	p
(a)	INR, per 1 unit increase	1.1 (0.73; 1.66)	0.6369
(b)	INR<2	1.28 (0.57; 2.84)	0.5492
	2≤INR≤3	Ref	-
	INR>3	1.95 (0.74; 5.18)	0.1792
(c)	TTR, per 10% increase	0.91 (0.76; 1.08)	0.2797
(d)	TTR ≥65% vs <65%	0.47 (0.2; 1.13)	0.0908
(e)	SD(INR), per 1 unit increase	2.44 (1.43; 4.15)	0.001
Model *	Predictors	Recurrent bleedings Intention to treat analysis § 55 events
		Hazard ratio (95% CI)	p
(a)	INR, per 1 unit increase	1.15 (0.83; 1.59)	0.4029
(b)	INR<2	1.15 (0.58; 2.28)	0.6889
	2≤INR≤3	Reference values	-
	INR>3	2.6 (0.88; 7.63)	0.0823
(c)	TTR, per 10% increase	0.88 (0.8; 0.95)	0.0023
(d)	TTR ≥65% vs <65%	0.35 (0.15; 0.8)	0.0129
(e)	SD(INR), per 1 unit increase	2.85 (1.71; 4.75)	0.0001

^*Cox model stratified on age (<75 years), co-variates included in the models: dialytic age, heart failure.

^§Statistically significant co-variates: (e) first bleeding=dialytic age (p=0.04459).

TTR was inversely associated with the risk of having recurrent haemorrhagic events (HR 0.88, 95% CI: 0.80; 0.95, p=0.0023) for each 10% increase. This association lost statistical significance when censoring at VKA withdrawal because of the lack of linearity (Figure 2B). When we compared TTR ≥65% vs <65% there was an over 50% reduction in the hazard of recurrent bleeding, both in the main and in the “as treated” analysis. SD(INR) was the parameter most strongly associated with both a first episode of bleeding and with recurrent bleedings, with an almost 3-fold increase in hazard ratio for each unit increment of SD(INR). This result was confirmed by the “as treated” analysis (Online Supplementary Table SII).

DISCUSSION

In the present study, the association between different indicators of oral anticoagulant therapy quality and the considered outcomes, i.e., mortality, first haemorrhagic episode and bleeding recurrences, were evaluated.

Variability of INR, as measured by its standard deviation, was the indicator most strongly related with the risk of both mortality and bleeding; therefore, INR variability had a clearly higher predictive value on survival than the INR value per se. This result suggests that it could be more dangerous for patients to have INR values ranging between over and under dosage, rather than being constantly over or under dosed. This is a clinically important concept, as it may suggest that the attempt to continuously adjust the VKA dosage when an out-of-range INR is found can be harmful rather than representing an advantage. This finding is confirmed by the fact that, as already seen in the general population, the relationship between INR values and mortality has a U-shape¹⁵. Recent studies using statistical approaches to mitigate the selection bias linked to prescription or non-prescription of VKAs (“propensity score” and “marginal structural models”) have shown that HD patients with AF taking VKAs are at lower risk of death compared to those who do not receive anticoagulants, especially in the presence of a high TTR^7,9,16–20. Our data suggest that also the variability of INR is a good predictor of survival in HD patients treated with VKAs. Our results are partly in agreement with those obtained by Sanden et al.²¹ in a large population of patients taking warfarin due to a previous episode of venous thromboembolism. In this study, both a low TTR and a high INR variability were associated with an increased risk of mortality. Furthermore, a Swedish study performed in a population of AF patients showed that INR variability, as assessed by the standard deviation of transformed INR, was a better predictor of mortality than TTR (HR of mortality 1.59 vs 1.18 for one standard deviation lower of TTR). When both INR variability and TTR were included in the same analysis, the influence of TTR on mortality risk was no longer significant¹³.

Bleeding data are also interesting. Again, the variability of INR proves to be the best indicator. A variation of 1 standard deviation of INR is associated with an at least 3-fold increase in both the first episode of bleeding and the recurrence of haemorrhage. TTR is strongly associated with bleeding recurrences, but not with the first bleeding event. These data suggest that, in HD patients with AF and high bleeding risk, it may not be sufficient to measure TTR, but it is also important to pay attention to the variability of the INR. The Swedish study mentioned above also showed that the variability of INR was more closely associated with the onset of bleeding than TTR and that TTR lost its predictive value when inserted into the same statistical model together with the SD(INR)¹³. Very few data regarding oral anticoagulant therapy indicators in CKD patients are available. Limdi et al. showed how CKD patients frequently receive an overdose of warfarin, resulting in an increased risk of bleeding⁵. Furthermore, a progressive worsening of the TTR was observed when the eGFR values decreased, with the worst performances in patients with severe CKD and undergoing HD⁵. As already mentioned, HD patients with AF have, in most cases, a clear indication for oral anticoagulant therapy, but unfortunately these patients also have a prohibitive risk of bleeding. Since the net clinical benefit of VKA treatment in HD patients is uncertain, and the quality of anticoagulant treatment is often poor, we looked for a tool which may help nephrologists take the difficult decision as to whether to continue VKA treatment or to stop it and look for different strategies, such as other antithrombotic drugs, left atrial appendage closure²², or even no treatment.

TTR was not designed to assess the quality of treatment in a single patient, but to find the right therapeutic range in different categories of patients¹². Parameters other than INR-TTR are under investigation to overcome this limitation²³. Moreover, electronic warfarin nomograms to guide dosing could improve INR control in HD patients, as already shown in the general AF outpatient setting²⁴. However, considering that it is very difficult to get a good TTR and stable INR in HD patients, DOACs may represent an attractive alternative to VKAs in these patients, since they do not need close monitoring by laboratory quality indicators. Data from the observational studies on the use of DOACs in dialysis patients do, however, differ^25–27, in particular regarding the risk of bleeding²⁸. For the moment, the guidelines do not recommend the use of these drugs in the dialysis population, as there are no randomized controlled trials that demonstrate their efficacy and safety. Thus, the tools currently available for the best possible oral anticoagulant therapy monitoring are extremely useful for these fragile patients.

CONCLUSIONS

Starting and managing VKA treatment in HD patients remains difficult due to the high risk of bleeding and the uncertainty of the benefit: risk ratio of anticoagulation. From our study, INR variability seems to be a reliable predictor of mortality, and bleeding episodes in HD patients taking VKAs and this parameter may help to estimate the net clinical benefit of the ongoing anticoagulant treatment for the patient. Moreover, our findings suggest that, in this population, it is better to consider more than one treatment quality indicator to identify patients at particularly high risk of mortality. However, the results of the study are limited because of the low number of enrolled patients. A larger multicentre study should be planned to verify our findings.