Predicting Stroke in Heart Failure and Preserved Ejection Fraction Without Atrial Fibrillation

BACKGROUND: The rate of stroke in patients with heart failure (HF) and preserved ejection fraction but without atrial fibrillation (AF), is uncertain as is whether it is possible to reliably predict the risk of stroke in these patients. METHODS: We validated a previously developed simple risk model for stroke among patients enrolled in the I-Preserve trial (Irbesartan in Heart Failure With Preserved Systolic Function) and PARAGON-HF trial (Efficacy and Safety of LCZ696 Compared to Valsartan, on Morbidity and Mortality in Heart Failure Patients With Preserved Ejection Fraction). The risk model consisted of 3 variables: history of previous stroke, insulin-treated diabetes, and plasma N-terminal pro-B-type natriuretic peptide level. RESULTS: Of the 8924 patients included in the pooled trial dataset, 5126 patients did not have AF at baseline. Among patients without AF, 190 (3.7%) experienced a stroke over a median follow-up of 3.6 years (rate 10.5 per 1000 patient-years). The risk for stroke increased with increasing risk score: second tertile hazard ratio, 1.78 (95% CI, 1.17–2.71); third tertile hazard ratio, 3.03 (95% CI, 2.06–4.47), with the first tertile as reference. For patients in the third tertile, the occurrence rate of stroke was 17.7 per 1000 patient-years, similar to that in patients with AF not receiving anticoagulation (20.7 per 1000 patient-years), and those with AF who were receiving anticoagulation (14.5 per 1000 patient-years). Model discrimination was good with a C index of 0.81 (0.68–0.91) and a simple score could be created from the model. CONCLUSIONS: A simple risk model can detect a subset of HF and preserved ejection fraction patients without AF who have a higher risk for stroke. The balance of risk-to-benefit in these individuals may justify the use of prophylactic anticoagulation, but this hypothesis needs to be prospectively evaluated. REGISTRATION: URL: https://www.clinicaltrials.gov; Unique identifiers: NCT00095238 and NCT01920711.


METHODS
The data that support the findings of this study are available from the corresponding author upon reasonable request.

Study Patients
Data from I-Preserve and PARAGON-HF were pooled to have a sufficient number of patients with HFpEF without AF for analysis. 16,17 Each trial was approved by local Ethics Committees and written informed consent was obtained from each patient. The design and primary results of I-Preserve and PARAGON-HF are already published. [16][17][18] Briefly, I-Preserve included 4128 patients at least 60 years of age and with a left ventricular ejection fraction (LVEF) of ≥45%. 16,18 Patients were required to have been hospitalized for HF during the previous 6 months and have New Your Heart Association (NYHA) class II to IV with corroborative evidence. Alternatively, if they had not been hospitalized, they were required to have NYHA class III to IV with corroborative evidence: chest X-ray (pulmonary congestion), electrocardiography (left ventricular hypertrophy, left bundle branch block), or echocardiogram (left ventricular hypertrophy, enlarged left atrium). NT-proBNP was not an inclusion criterion but NT-proBNP was measured at baseline in most patients (although the assay results were not known to investigators). Patients were randomly assigned in a 1:1 ratio to receive irbesartan 75 mg once daily (target dose 300 mg) or a matching placebo. The median follow-up was 49.5 months.
WHAT IS NEW?
• The rate of occurrence of stroke in patients with heart failure and preserved ejection fraction but without atrial fibrillation is unclear, as is how to reliably predict the risk of stroke in such patients. We validated the previously developed stroke risk model consisting of 3 variables: previous stroke history, insulin-treated diabetes, and N-terminal pro-B-type natriuretic peptide level. Patients in the highest tertile of this risk model had 3 times the risk of stroke compared to the lowest tertile.

WHAT ARE THE CLINICAL IMPLICATIONS?
• This risk model reliably identified a subset of patients with heart failure and preserved ejection fraction without atrial fibrillation at a high risk of stroke. The model can be converted to a simple score (S 2 I 2 N 0-3 ) convenient for clinical use. • Patients at high risk of stroke may have a risk-benefit balance that justifies the use of prophylactic anticoagulation, although this needs to be tested prospectively in a clinical trial.  17 Patients were required to have signs and symptoms of HF, NYHA class II to IV, evidence of structural heart disease, and diuretic therapy. For patients who were hospitalized for HF within 9 months, those in AF on screening electrocardiography were required to have an NT-proBNP concentration ≥600 pg/mL, and those not in AF were required to have an NT-proBNP concentration ≥200 pg/mL. For patients without hospitalization for HF within 9 months, those in AF on screening electrocardiography were required to have an NT-proBNP concentration ≥900 pg/mL, and those not in AF were required to have an NT-proBNP concentration ≥300 pg/ mL. Patients entered a single-blind run-in period of 1 to 2 weeks of treatment with valsartan 40 or 80 mg twice daily followed by a period of 2 to 4 weeks of treatment with sacubitril/valsartan 49/51 mg twice daily. Thereafter, patients were randomly assigned in a 1:1 ratio to double-blind treatment with either sacubitril/valsartan 97/103 mg twice daily or matching valsartan 160 mg twice daily. The median follow-up was 35 months.

Stroke Diagnosis
The occurrence of stroke was a secondary end point and was centrally adjudicated by a clinical events committee in both trials. [16][17][18] Stroke in both trials was defined as a focal neurological deficit of central origin lasting more than 24 hours (except for death within 24 hours), with or without imaging confirmation of cerebral infarction or intracerebral hemorrhage. However, categorization of stroke by cause (ischemic, hemorrhagic, or other) was only available in the PARAGON-HF trial.

New-Onset of AF
The new onset of AF was prospectively collected using a specific case report form in I-Preserve. 16,18 The new occurrence of AF was a prespecified secondary end point in PARAGON-HF. 17 However, there was no systematic ECG surveillance for AF in either trial.

Statistical Methods
Patients with AF were defined as those with either AF confirmed on their baseline ECG or a prior history of AF and the remaining patients were defined as those without AF. Data regarding AF on ECG and a prior history of AF were missing in 18 cases and 4 cases, respectively, in PARAGON-HF. Descriptive statistics were used to describe the whole cohort and to compare these 2 subgroups, with means±SD, medians (interquartile range) for continuous variables, or number (percentage) for categorical variables. We also compared the baseline characteristics of patients who developed stroke during the trial and those without.
The rate of occurrence of stroke (per 1000 patient-years) was calculated during the trial follow-up period and compared between the aforementioned subgroups. Cumulative incidence function (CIF) plots were drawn for survival analyses. We estimated CIF for stroke occurrence considering the competing risk of death. To meet the assumption of the independence of stroke events, the first event in a patient after randomization was evaluated in the analysis.
We applied the previously published risk model for stroke occurrence derived from CORONA/GISSI-HF to the pooled data in patients without AF from the HFpEF trials. 14 The risk score was calculated by the following equation: (history of a previous stroke)×6.53+(insulin-treated diabetes)×7.39+(plasma NT-proBNP measurement at baseline [pmol/L] [in logarithmic transformation])×2.80. NT-proBNP units pg/mL were converted to pmol/l, with 1 pg/mL=0.1182 pmol/L. One-year, 2-year, and 3-year rates of occurrence were estimated by the following equation: 1-year, 1−0.9971^exp(risk score/10); 2-year, 1−0.9945^exp(risk score/10); 3-year, 1−0.9908^exp(risk score/10). 14,15 As transient ischemic attack and stroke history were not collected separately in I-Preserve, the risk score was calculated by considering transient ischemic attack or stroke history as stroke history. Patients with a missing value for a history of previous stroke (n=4), insulin-treated diabetes (n=1), and NT-proBNP (n=457) were excluded from the model calculation, and complete case analyses were performed for the evaluation of the model and estimation of the rate of occurrence. Cox proportional hazard model was conducted to compute the hazard ratios (HR) and 95% CI of the tertiles of the risk score. According to the tertiles, CIF plots for stroke occurrence were obtained. We evaluated the model discrimination using the overall C index for the risk model according to the method of Pencina and D' Agostino, as outlined by Liu et al. 19,20 We also assessed the C statistics of the model, using the traditional Harrell C statistic. 21 The calibration of the model and its ability to separate the patient population into risk groups were assessed by observing the predicted and observed outcomes in the tertiles. Finally, discrimination of the S 2 I 2 N 0-3 score, which we previously proposed based on the aforementioned risk model, was evaluated. 15 To examine the association between a stroke and subsequent mortality, Kaplan-Meier curves were plotted. At baseline, all patients were in the no stroke group and changed exposure to stroke after a first stroke (or stayed in the no stroke group). The hazard ratio (and 95% CI) for mortality after a stroke (with the no stroke group as a reference), adjusted for age, sex, NYHA functional class, body mass index, systolic blood pressure, heart rate, serum creatinine, NT-proBNP, coronary heart disease, diabetes, and history of stroke, was computed using the Cox proportional hazard models.

RESULTS
Of the 4128 patients in I-Preserve, 1233 (29.9%) had either a history of AF or AF on their baseline ECG. The corresponding number was 2565 (53.5%) of the 4796 patients in PARAGON-HF. This generated a total of 3798 patients (42.6%) with AF and 5126 patients without AF in the pooled dataset.

Patients With and Without AF
The baseline demographics of patients with and without AF are shown in Table S1. Patients without AF were younger and more often female and had a higher LVEF and worse NYHA functional class. Levels of serum creatinine and plasma NT-proBNP were lower in patients without AF than in those with AF. Regarding comorbidity, a history of coronary heart disease was more commonly observed in patients without AF compared with those with AF. The 2 groups had a similar prevalence of diabetes, but those without AF were treated with insulin for their diabetes more frequently and more often had a history of stroke. A beta-blocker, mineralocorticoid antagonist, and digoxin were less frequently prescribed for patients without AF, and notable differences were observed in the use of antiplatelet therapy (40.3% of patients without AF versus 21.3% in those with AF) and anticoagulant treatment (10.9% versus 58.6%, respectively). Table 1 shows the baseline characteristics of patients without AF, according to whether or not patients developed a stroke after randomization. Patients without AF who experienced a stroke were slightly older than those who did not, but the proportion of females was similar in the 2 groups. Patients who developed a stroke were more often of Black race, had higher blood pressure at baseline, and higher creatinine and NT-proBNP levels than those who did not. A history of prior stroke was more common in patients who developed a stroke during follow-up. There was also a trend for more insulin-treated diabetes in patients who developed a stroke during follow-up. The baseline characteristics according to the occurrence of stroke in patients with and without AF are shown in Table S2.

Rates of Stroke
The median follow-up in the pooled analysis was 3.4 years and 396 (4.4%) patients developed a stroke (13.1 per 1000 patient-years). In I-Preserve, 196 patients developed a stroke (11.9 per 1000 patient-years), and in PARAGON-HF, 200 patients had a stroke (14.7 per 1000 patient-years).

Validation of the Stroke Prediction Model in Patients Without AF
The distribution of the stroke risk score is shown in Figure S1. The median value of the risk score was 11.4, and when patients were classified into 3 equally sized groups according to their risk score, it was 7.8 in tertile 1, 11.4 in tertile 2, and 17.0 in tertile 3. The CIF plots for stroke according to the tertile of the risk score are shown in Figure 2. The numbers of strokes in tertiles 1, 2, and 3 were 38, 53, and 83 respectively. The 1-, 2-and 3-year CIF rates of stroke in the highest tertile were 1.8% (95% CI, 1.3%-2.6%), 3.4% (95% CI, 2.6%-4.5%), and 4.6% (95% CI, 3.7%-5.8%), respectively. Patients in risk-tertile 3 had an overall stroke rate of 17.7 per 1000 patient-years. In Cox proportional hazard models, the risk of stroke increased as the risk score increased (

Model Calibration and Discrimination
Observed and predicted probabilities of a stroke at 1, 2, and 3 years were compared with the patients divided by tertiles ( Figure 3) and were acceptable. HRs according to tertiles were similar even when we took into account allocated treatment (Table S3). Model discrimination was good: the overall C index was 0.81 (95% CI, 0.68-0.91). The Harrell C statistic is available in the Supplemental Material (Table S4).

S2I2N0-3 Score
The number of patients, strokes observed, and the predicted incidence of stroke at 1 year according to S 2 I 2 N 0-3 score are shown in Tables 3 and 4. The score  Data are presented as mean±SD, median (interquartile range), or number (percentage). NYHA class was missing in 2 cases, time from diagnosis of HF 11 cases, LV ejection fraction 1 case, body mass index 8 cases, heart rate 1 case, serum creatinine 56 cases, NT-proBNP 457 cases, history of stroke 4 patients, current smoker 19 patients, and medical history including insulin 1 patient. AF indicates atrial fibrillation; CABG, coronary artery bypass graft; HF, heart failure; LV, left ventricular; NT-proBNP, N-terminal pro-B-type natriuretic peptide; NYHA, New York Heart Association; and PCI, percutaneous coronary intervention.

Association Between a Stroke and Subsequent Mortality
In participants without AF, compared to patients with no stroke, the risk of death markedly increased after a stroke: all-cause mortality rate 4.0 (95% CI, 3.7-4.3) per 100 patient-years in patients with no stroke versus 27.8 (95% CI, 22.1-35.0) per 100 patient-years in patients after a stroke -giving an HR of 6.90 (95% CI, 5.32-8.95; Figure S2). The difference in risk of death was large over the initial 30 days after a stroke but remained significant beyond 30 days.

DISCUSSION
In the present study, we confirmed that a simple model consisting of 2 clinical variables (history of previous stroke and insulin-treated diabetes) and a routinely measured biomarker (NT-proBNP) successfully predicted the stroke risk in patients with HFpEF without AF; the discrimination of this model for stroke risk was good and the predictive probability was accurate. The rate of occurrence of stroke among patients without AF in the highest tertile of risk (17.7 per 1000 patient-years) was close to that of individuals with AF and not treated with an anticoagulant (20.7 per 1000 patient-years) and higher than in those with AF who were treated with an anticoagulant (14.5 per 1000 patient-years). Few strokes were preceded by clinically recognized AF. Finally, the risk of death increased considerably after a stroke.
Little epidemiological information on the occurrence of stroke in patients with HFpEF without AF is available. 3 One of the few such sources is the Swedish Heart Failure Registry, which showed that the rate of occurrence of ischemic stroke or transient ischemic attack in these patients was 17.9 per 1000 patient-years, which is considerably higher than the rate in our study (10.5 per 1000 patient-years). 11 However, the Swedish population was older, had higher NT-proBNP levels, and included some individuals with an LVEF between 40 and 45%; the composite outcome also included transient ischemic attack, and all of these may explain the different event rates. Two recent reports have provided the rate of stroke in the TOPCAT trial (Treatment of Preserved Cardiac Function Heart Failure With an Aldosterone Antagonist) but did not differentiate between patients with and without AF. 22,23 Therefore we analyzed the TOPCAT dataset (Americas only) to differentiate between patients with and without AF. Among the 1007 patients not in AF at baseline, 36 strokes occurred during a median follow-up of 2.6 years, giving a stroke rate of 12.4 (9.0-17.3) per 1000 patientyears, consistent with our findings (and lower than the rate among the 760 patients in TOPCAT with AF in whom the stroke rate was 18.7, 13.8-25.3, per 1000 patient-years).
A comparison with patients with HFrEF who do not have AF is also of interest. In a recent analysis using a  Heart Failure]), we observed a rate of stroke of 11.7 per 1000 patient-years. 15,[24][25][26] The similar stroke rate in the 2 major HF phenotypes is notable given the previously reported relationship between LVEF and stroke occurrence. 27 Prior concepts of blood stasis associated with reduced left ventricular contractility leading to thrombosis and embolism may be too simplistic and do not explain the similar rate of stroke in HFrEF and HFpEF. Prior stroke is expected to be predictive of future stroke and type 2 diabetes requiring insulin is usually longstanding and often associated with widespread endothelial dysfunction, atherosclerosis, and abnormalities of coagulation and fibrinolysis, as well as nephropathy, all of which are associated with a higher risk of stroke. [28][29][30] The association with higher NT-proBNP is perhaps less obvious, but this may reflect atrial enlargement/myopathy and even occult paroxysmal AF. [31][32][33][34] The obvious therapeutic question raised by our findings is whether the risk of stroke in patients with HFpEF can be reduced. Specifically, might anticoagulation play such a role? In WARCEF, the risk of ischemic stroke was reduced by almost half with warfarin (29 versus 55 strokes; hazard ratio, 0.52 [95% CI, 0.33-0.82]) in patients with HFrEF in sinus rhythm; however, there was a small excess of intracerebral hemorrhage (5 versus 2 cases). In similar patients in COMMANDER-HF, the number of ischemic strokes was smaller in the rivaroxaban group compared with the placebo group (41 versus 63; hazard ratio, 0.64 [95% CI, 0.43-0.95]), but there was no excess of intracranial bleeding. A recent systematic review of randomized controlled clinical trials assessing oral anticoagulants versus placebo or antiplatelet agents in patients with HF or ventricular systolic dysfunction/cardiomyopathy without clinical HF, and sinus rhythm found a total of 7 trials which included 15 794 patients. 35 In that report, oral anticoagulation reduced the rate of stroke or systemic embolism compared to control (odds ratio, 0.57 [95% CI, 0.39-0.82]). Collectively, these reports suggest an important role for thrombosis or thromboembolism in the causation of stroke, at least in HFrEF. However, in unselected patients, the benefit-to-risk balance is not sufficiently favorable to recommend treating all patients with HFrEF with an anticoagulant. For example, in COMMANDER-HF, there were 8 more patients per 1000 patient-years of treatment with major bleeding but only 5 fewer patients with stroke per 1000 patient-years of treatment with rivaroxaban, compared with placebo. Hence, we have argued that anticoagulation should be targeted at patients at the highest risk of stroke, assuming such patients can be easily and reliably identified. We think that our prediction model fulfills this goal, now having been validated in both HFpEF and HFrEF. The simple S 2 I 2 N 0-3 score we have created enables this model to be used easily in clinical practice. Since the components of this score change over time, it may be appropriate to reassess the score during a patient's follow-up. The recent emergence of factor XI inhibitors potentially strengthens the approach we have suggested because these novel agents seem to carry a very low risk of bleeding and targeted to patients at higher risk of stroke may further tip the benefit-to-risk balance in a favorable direction. [36][37][38][39] Although questions remain regarding the relative roles of thrombosis and thromboembolism in HFpEF versus HFrEF, in patients with AF, anticoagulant therapy is equally effective in individuals with HFrEF and HFpEF. 32 Clearly, this   hypothetical strategy of stroke risk-stratification and targeted anticoagulation needs to be tested in a prospective randomized controlled trial. 3,40 Finally, our results emphasize the importance of primary and secondary prevention of stroke given not only the disability that results from this event but also because of the greatly elevated risk of death occurring after stroke. Since several risk factors for the development of stroke, including diabetes and hypertension, also risk factors for the development of HFpEF (and comorbidities targeted in the management of HFpEF), these deserve special attention in the older population at risk of stroke and HFpEF as well as in people who have developed HFpEF or experienced stroke. 1,6,8,11,14 There are several limitations to our study. First, the 2 large clinical trials used in our analyses have specific inclusion/exclusion criteria and likely included patients at lower risk than in the real world, including lower risk of stroke (eg, patients with prior disabling stroke may not have been enrolled). 11 Second, we could not distinguish between type 1 and type 2 diabetes although the majority of patients with HFpEF have type 2 rather than type 1 diabetes. Third, although the new occurrence of AF was collected prospectively in each study, systematic electrocardiographic monitoring was not performed. Thus, the reported incidence of AF is likely lower than would have been detected by electrocardiographic monitoring. However, screening for AF is currently not recommended or feasible for all patients with HFpEF. In any case, because a stroke may occur at the time of or shortly after the onset of AF, an AF-detection strategy is likely to be less effective at reducing the risk of stroke than prophylactic anticoagulation. Finally, we could not differentiate between ischemic and hemorrhagic stroke, although hemorrhagic stroke is thought to be relatively uncommon, as has been shown in the HFrEF anticoagulation trials. 11 Moreover, in our previous HFrEF analyses, the risk model described here was as effective at predicting the occurrence of ischaemic stroke as overall stroke. 15 In conclusion, we confirmed that patients with HFpEF can have a substantial risk of stroke even in the absence of AF and validated a risk model for stroke in patients with HFpEF without AF. This simple risk model can detect a subset of patients with HFpEF without AF who have a high rate of occurrence of stroke. The balance of risk-to-benefit in these individuals may justify the use of prophylactic anticoagulation. This hypothesis needs to be evaluated in a prospective randomized controlled trial.

Disclosures
Dr Kondo received speaker fees from Abbott, Ono Pharma, Otsuka Pharma, Novartis, AstraZeneca, Bristol Myers Squibb, and Abiomed. Dr Jering is supported by National Heart, Lung, and Blood Institute T32 postdoctoral training grant T32HL007604. Dr Jhund's employer the University of Glasgow has been remunerated by AstraZeneca, personal fees from Novartis and Cytokinetics, and grants from Boehringer Ingelheim. Dr Anand reports receiving fees for serving on a steering committee from AstraZeneca, ARCA biopharma, Amgen, and Liva-Nova; fees for serving as chair of a data and safety monitoring board from Boston Scientific; fees for serving on an end point committee from Boehringer Ingelheim; and fees for serving on an advisory board from Zensun. Dr Desai has received grants and personal fees from AstraZeneca during the conduct of the study; personal fees from Abbott, Biofourmis, Boston Scientific, Boehringer Ingelheim, Corvidia, DalCor Pharma, Relypsa, Regeneron, and Merck; grants and personal fees from Alnylam and Novartis; and personal fees from Amgen, outside the submitted work. Dr Lam is supported by a Clinician Scientist Award from the National Medical Research Council of Singapore; has received research support from AstraZeneca, Bayer, Boston Scientific, and Roche Diagnostics; has served as a consultant or on the advisory board/steering committee/executive committee for Actelion, Amgen, Applied Therapeutics, AstraZeneca, Bayer, Boehringer