The novel risk score model for predicting the poor anticoagulation control in patients with atrial fibrillation taking warfarin

Ischemic stroke is the most catastrophic sequelae in patients with atrial fibrillation (AF). Oral anticoagulants (OACs) are recommended by many international clinical practice guidelines in these patients, except in patients with very low-risk defined as CHA₂DS₂-VASc score of 0 in males and 1 in females [1–4]. Warfarin is the common OAC used in AF patients in Thailand [5]. Previous studies have shown that the time in therapeutic range (TTR) that targets the international normalized ratio (INR) between 2 and 3 of AF patients taking warfarin was 46.22%–53.17% [6, 7]. Moreover, previous trials have demonstrated that poor TTR is associated with adverse events including thromboembolism, bleeding, and/or mortality [8–10].

The SAMe-TT₂R₂ score was studied to predict the suboptimal TTR in patients taking vitamin K antagonists (VKAs) [7, 11–17]. However, the C-statistics of the SAMe-TT₂R₂ score in the prediction of suboptimal TTR is very low [17, 18]. Moreover, the “non-Caucasian” race that had 2 points on the SAMe-TT₂R₂ score limited the use of this score in Asians. To date, the European guidelines and the Asia-Pacific Heart Rhythm Society have recommended to consider VKA or non-vitamin K antagonist oral anticoagulants (NOACs) if the SAMe-TT₂R₂ score is 0–2 [1, 3]. However, a recent trial has demonstrated that AF patients taking warfarin have a significantly higher cardiovascular death or ischemic stroke/transient ischemic attack/systemic embolism or major bleeding compared with those taking NOACs, regardless of the SAMe-TT₂R₂ score [19]. This study aimed to propose the development of a novel risk score model for predicting the poor anticoagulation control in AF patients taking warfarin and compare the new risk score model with the SAMe-TT₂R₂ score.

We prospectively recruited AF patients aged ≥18 years from 27 hospitals between 2014 and 2017 in the COhort of antithrombotic use and Optimal INR Level in patients with non-valvular AF in Thailand (COOL AF Thailand) registry. We excluded patients having prosthetic heart valve, rheumatic mitral valve disease, a history of recent ischemic stroke within 3 months, transient reversible cause of AF, predicted life expectancy <3 years, pregnancy, thrombocytopenia, and myeloproliferative diseases from this study.

We retrieved demographic and clinical data of AF patients taking warfarin stratified by anticoagulation control. We collected the baseline demographic data such as age, sex, body mass index (BMI), AF pattern and duration from AF diagnosis, AF-related symptoms, medical history including smoking and alcohol status, CHA₂DS₂-VASc score, HAS-BLED score, SAMe-TT₂R₂ score, left ventricular (LV) function, and concomitant medication. We recorded patient data at follow-up visits every 6 months. The poor anticoagulation control was defined as TTR <65%. We used the Rosendaal method for calculating TTR [20].

Each component of the SAMe-TT₂R₂ score was scored and recorded as S = female sex (1 point); A = age <60 years (1 point); Me = medical history >2 of the following: hypertension, diabetes, coronary artery disease (CAD)/myocardial infarction (MI), peripheral arterial disease (PAD), congestive heart failure (HF), previous stroke, pulmonary disease, and hepatic or renal disease (1 point); T = treatment (interacting drugs, e.g., amiodarone for rhythm control) (1 point); T₂= tobacco use within 2 years (2 points); and R₂ = non-Caucasian race (2 points).

This study protocol was approved by the Central Research Ethics Committee (CREC) (Certificate No.: COA-CREC003/2014). All participating patients provided written informed consent. We conducted this study protocol in compliance with the Declaration of Helsinki and the International Conference on Harmonization for Good Clinical Practice Guidelines (ICH-GCP).

Of 3,461 patients, 2,233 patients taking warfarin having available TTR data were retrieved into this study. There were 1,432 patients having poor anticoagulation control (TTR <65%) and 801 patients having good anticoagulation control (TTR ≥ 65%) (Figure 1). The average age was 68.9 ± 10.6 years, and 44% were females. Baseline characteristics of the patients are shown in Table 1.

Figure 1.

Flow diagram of study patients. INR, international normalized ratio; TTR, time in therapeutic range.

Univariate and multivariate analysis showed that symptomatic AF, diabetes, HF, and a history of bleeding were associated with increased risk while obesity, AF duration, and paroxysmal AF were associated with decreased risk of poor anticoagulation control (Table 2). Diabetes, HF, and a history of bleeding were the most significant predictors of poor anticoagulation control. We scored each acronym of these predictors and generated the novel risk score model. The SHOB-D₂AF score was created. The definition of this score is illustrated in Table 3. Mean TTR stratified by the SHOB-D₂AF score and SAMe-TT₂R₂ score are shown in Table 4. ROC curve analysis showed that a SHOB-D₂AF score of 4 or more could be used to predict poor anticoagulation control at the best cut-off point and predict better than the SAMe-TT₂R₂ score of 3 or more. The sensitivity, specificity, PPV, and NPV of the SHOB-D₂AF score in predicting poor anticoagulation control were 47.9% (95% confidence interval [CI] 45.3%–50.5%), 63.3 (95% CI 59.9%–66.6%), 70.0% (95% CI 67.1%–72.8%), and 40.5% (95% CI 37.8%–43.2%), respectively, while the sensitivity, specificity, PPV, and NPV of the SAMe-TT₂R₂ score in predicting poor anticoagulation control were 76.3% (95% CI 74.0%–78.4%), 23.1 (95% CI 20.3%–26.2%), 64.0% (95% CI 61.7%–66.2%), and 35.2% (95% CI 31.3%–39.4%), respectively (Table 5). The SHOB-D₂AF score significantly improved the discrimination performance compared with the SAMe-TT₂R₂ score (C-statistics 0.584 [0.563–0.604] vs 0.506 [0.485–0.527], P < 0.001) (Figure 2A). Sensitivity analysis was performed for comparison of the ROC curve from predicted probability derived from the SHOB-D₂AF score and the SAMe-TT₂R₂ score. The C-statistics of predicted probability of the SHOB-D₂AF score was comparable to the SAMe-TT₂R₂ score (0.586 [0.565–0.607] vs 0.572 [0.551–0.592], P = 0.255) (Figure 2B). The NRI was 17.82% (9.20%–26.44%), indicating that the SHOB-D₂AF score performed better than the SAMe-TT₂R₂ score in predicting poor anticoagulation control (P< 0.001). The IDI was 1.80% (1.06%–2.54%) (P < 0.001) (Figure 3).

Figure 2.

ROC curve illustrating discrimination performance between (A) SHOB-D₂AF score and SAMe-TT₂R₂ score and (B) predicted probability derived from SHOB-D₂AF score and SAMe-TT₂R₂ score. AF, atrial fibrillation; AUC, area under the curve; ROC, receiver operating characteristic.

Figure 3.

NRI of SHOB-D₂AF compared with SAMe-TT₂R₂ model. AF, atrial fibrillation; IDI, Integrated Discrimination Index; NRI, Net Reclassification Index; TTR, time in therapeutic range.

To our knowledge, high TTR is the most important marker to consider warfarin as an option for OAC in AF patients. Additionally, poor TTR has been associated with adverse outcomes in these patients [8–10]. The SHOB-D₂AF score is the novel risk score model for predicting poor anticoagulation control in AF patients taking warfarin. The TTR decreased following the increased scores of these risk score models. Of note, the SHOB-D₂AF score could be used to predict poor TTR better than the SAMe-TT₂R₂ score in this cohort study in terms of area under the curve (AUC) and NRI.

Previous trials have shown that the SAMe-TT₂R₂ score could predict the poor anticoagulation control in AF patients taking VKAs [7, 11–17]. However, those trials demonstrated that there were variable results of suboptimal INR prediction in terms of AUC from 0.54 to 0.70 [7, 11, 12, 16, 17]. In Thailand, previous SAMe-TT₂R₂ score trials have shown that AUC was 0.54–0.59, respectively [7, 10, 17]. Nevertheless, recent European guidelines recommend to choose VKAs or NOACs for stroke prevention stratified by the SAMe-TT₂R₂ score [1]. A previous study revealed that the SAMe-TT₂R₂ score could not be used to choose VKAs or NOACs in AF patients taking warfarin [19].

The SAMe-TT₂R₂ score having lower than expected accuracy for prediction of poor anticoagulation control may be from the magnitude of predictability in each parameter in this risk score. For example, there are many illnesses in medical history such as hypertension, diabetes, CAD/MI, PAD, congestive HF, previous stroke, pulmonary disease, and hepatic or renal disease. Each medical illness has a different predictability for poor anticoagulation control, such as congestive HF may have stronger predictability than hypertension or CAD/MI. This study showed that only diabetes and HF were the predictors of poor anticoagulation control, while hypertension, CAD/MI, PAD, previous stroke, CKD, and liver cirrhosis were not the predictors. Moreover, other risk predictors of poor anticoagulation control did not mention in the SAMe-TT₂R₂ score symptoms such as BMI, duration from AF diagnosis, pattern of AF, a history of bleeding, or AF-related symptoms. This trial demonstrated that these risk factors could predict poor anticoagulation control as well. Besides, the “non-Caucasian” race item in the SAMe-TT₂R₂ score makes this scoring system not suitable to be used in the Asian population. Finally, this study showed that the SHOB-D₂AF score had superior PPV with comparable NPV and significantly improved the prediction of poor anticoagulation control in terms of AUC and NRI compared with the SAMe-TT₂R₂ score. In clinical practice, AF patients with a SHOB-D₂AF score ≥4 should be recommended to choose NOACs over VKAs, while those patients with a SHOB-D₂AF of <4 should be recommended to initiate VKAs first as well.

However, this study has several limitations. First, there were small patients having some risk predictors for poor anticoagulation control such as PAD, liver cirrhosis, dementia, hypothyroidism, hyperthyroidism, and amiodarone. However, our risk score model still had more predictors and could predict better poor anticoagulation than the SAMe-TT₂R₂ score. Second, this study was an observational study leading to selection bias. Nevertheless, we analyzed all predictors in univariate and multivariate analysis. However, our risk score model needs validation in another dataset of Thai patients. Finally, our patients were Thai, leading to limit generalizability. The non-Caucasian race is the predictor of poor anticoagulation control in the SAMe-TT₂R₂ score, but this study has had no other race included in our analysis. Our risk score model will need to conduct studies on other races in the future.

The SHOB-D₂AF score was the novel risk score model that was better than the SAMe-TT₂R₂ score in predicting poor anticoagulation control.