Comparison of the Performance of the Warfarin Pharmacogenetics Algorithms in Patients with Surgery of Heart Valve Replacement and Heart Valvuloplasty

Thrombosis Research 136 (2015) 552–559

Contents lists available at ScienceDirect

Thrombosis Research journal homepage: www.elsevier.com/locate/thromres

Full Length Article

Comparison of the Performance of the Warfarin Pharmacogenetics Algorithms in Patients with Surgery of Heart Valve Replacement and Heart Valvuloplasty Hang Xu a,1, Shi Su b,1, Wuji Tang b, Meng Wei c, Tao Wang d, Dongjin Wang d,⁎, Weihong Ge a,⁎⁎ a

Department of Pharmacy, Affiliated Drum Tower Hospital of Nanjing University Medical School, Nanjing, Jiangsu, China Department of Clinical Pharmacy, China Pharmaceutical University, Nanjing, Jiangsu, China c Department of Pharmacy, Affiliated Jinling Hospital of Nanjing University Medical School, Nanjing, Jiangsu, China d Department of Cardiothoracic Surgery, Affiliated Drum Tower Hospital of Nanjing University Medical School, Nanjing, Jiangsu, China b

a r t i c l e

i n f o

Article history: Received 27 February 2015 Received in revised form 12 June 2015 Accepted 30 June 2015 Available online 6 July 2015 Keywords: Warfarin Pharmacogenetics Dosing algorithm Performance

a b s t r a c t A large number of warfarin pharmacogenetics algorithms have been published. Our research was aimed to evaluate the performance of the selected pharmacogenetic algorithms in patients with surgery of heart valve replacement and heart valvuloplasty during the phase of initial and stable anticoagulation treatment. 10 pharmacogenetic algorithms were selected by searching PubMed. We compared the performance of the selected algorithms in a cohort of 193 patients during the phase of initial and stable anticoagulation therapy. Predicted dose was compared to therapeutic dose by using a predicted dose percentage that falls within 20% threshold of the actual dose (percentage within 20%) and mean absolute error (MAE). The average warfarin dose for patients was 3.05 ± 1.23 mg/day for initial treatment and 3.45 ± 1.18 mg/day for stable treatment. The percentages of the predicted dose within 20% of the therapeutic dose were 44.0 ± 8.8% and 44.6 ± 9.7% for the initial and stable phases, respectively. The MAEs of the selected algorithms were 0.85 ± 0.18 mg/day and 0.93 ± 0.19 mg/day, respectively. All algorithms had better performance in the ideal group than in the low dose and high dose groups. The only exception is the Wadelius et al. algorithm, which had better performance in the high dose group. The algorithms had similar performance except for the Wadelius et al. and Miao et al. algorithms, which had poor accuracy in our study cohort. The Gage et al. algorithm had better performance in both phases of initial and stable treatment. Algorithms had relatively higher accuracy in the N50 years group of patients on the stable phase. © 2015 Elsevier Ltd. All rights reserved.

1. Introduction Warfarin, the most common oral anticoagulants, is widely prescribed for the prevention of thromboembolism disorders such as deep vein thrombosis, stroke, and pulmonary embolism. Though a lifelong medication for patients with the heart valve replacement surgery, warfarin has a narrow therapeutic index [1], which leads to the increased risk of adverse events like bleeding for overdose and thrombosis for insufficient treatment [2]. It is important for patients to get appropriate prescriptions of warfarin, especially for patients in the early phase of post-surgery who had surgery for heart valve ⁎ Correspondence to: D. Wang, Department of Cardiothoracic Surgery, Affiliated Drum Tower Hospital of Nanjing University Medical School, No.321 Zhongshan Rd, NanjingCity, Jiangsu Province, 210008, PR China. ⁎⁎ Correspondence to: W. Ge, Department of Pharmacy, Affiliated Drum Tower Hospital of Nanjing University Medical School, No.321 Zhongshan Rd, NanjingCity, Jiangsu Province, 210008, PR China. E-mail addresses: [email protected] (D. Wang), [email protected] (W. Ge). 1 Hang Xu and Shi Su contributed equally to the manuscript as co-first authors.

http://dx.doi.org/10.1016/j.thromres.2015.06.032 0049-3848/© 2015 Elsevier Ltd. All rights reserved.

replacement and heart valvuloplasty (HV). The risk for thromboembolism disorders can be decreased by shortening the time to reach the therapeutic INR (International Normalized Ratio) [3]. The metabolism of warfarin differs among individuals as a result of the single nucleotide polymorphisms (SNPs) in genes of cytochrome P450 (CYP) 2C9 and vitamin K epoxide reductase (VKORC1) [4–6]. Clinical factors such as race, age, gender, body weight, diet and concomitant medication also have relevance to the diversity of the warfain dose requirements [7–9]. A number of algorithms [10–19] were derived by the incorporation of gene factors and clinical factors. Theses algorithms were established to make warfarin prescriptions more accurate for individual patients. The algorithms were also aimed to shorten the time needed to reach the therapeutic INR and keep the anticoagulation treatment safe. The performance of some algorithms has been validated by several researches [20,21]. However, there is little data of verifications conducted in the cohort of patients undertaking mechanical heart valve replacement. Most of the algorithms in published papers were derived from patients in the steady state of anticoagulation. Sensitivity to warfarin

H. Xu et al. / Thrombosis Research 136 (2015) 552–559

increases in the patients after heart valve surgery [22–24], and a progressive decrease in sensitivity to warfarin has been discovered in these patients [25]. The performance of the algorithms is needed to be validated in periods of initial and stable anticoagulation treatment. The performance of warfarin dosing algorithms might be influenced by specific factors of derivation population, such as age and races. Since chronic heart failure has a high prevalence in patients undertaking heart surgery, it might also be the factors that contributed to the performance variation. In the study, we aimed to validate the performance of the established pharmacogenetics-based algorithms by a cohort of patients undertaking heart valve surgery during the period of initial and stable periods of anticoagulation therapy. Accuracy of pharmacogeneticsbased algorithms was also evaluated in subgroups divided by dose range, heart function and age. 2. Patients and Methods 2.1. Study Design The genotyping of CYP2C9*3 (1075G N T, rs1057910) and VKORC1 (1639G N A, rs9923231) was accomplished after the recruitment of the patients. Warfarin was prescribed the day after the surgery. The INR was aimed to control in the range of 1.8-2.5 and was measured every day for the initial anticoagulation therapy. The initial dose of wafarin was defined as the constant dose prescribed to obtain 3 consecutive therapeutic INRs at least 5 days after the anticoagulation therapy started. Follow-up data were collected by the anticoagulation clinic of the Drum Tower Hospital Affiliated to Nanjing University Medical School. Stable anticoagulation was indicated by patient received same dose of warfarin that led to 3 consecutive INR tests separated by at least 1 week within target range from 1 year after the surgery. 2.2. Patients A total of 212 Han-Chinese patients were registered consecutively from the Affiliated Drum Tower Hospital of Nanjing University from July 2012 to December 2012 for the prospective study. Patients were recruited by criteria as follows: (1) Be admitted to hospital for heart valve surgery and would need to receive warfarin for anticoagulation therapy after the surgery; (2) The patients had to be at least 18 years old;(3) Patients had the willingness to participate in the study and the ability to provide written informed consent; (4) Able to followed the schedule for anticoagulation visit and had adherence of warfarin prescription. All of these patients started warfarin therapy the day after their surgeries. Patients were excluded by criteria as follows: (1) Contraindications of anticoagulation or holding warfarin during the following-up period; (2) Patients with abnormal liver function or renal impairment that may influence the accuracy of the study; (3) Pregnancy or lactating; (4) Patients with abnormal function of coagulation due to liver disease, antiphospholipid antibody; (5) Poorly cognitive of informed consent or lack of compliance of prescription for warfarin and INR test; (6) Abuse of alcohol or irregular diet.The protocol of this study was approved by the Ethics Review Board of the Affiliated Drum Tower Hospital of Nanjing University. The protocol was admitted for the fulfillment of the Declaration of Helsinki. Patients were recruited with written informed consent. 2.3. Data Collection The demographic data of the patients, including gender, age, height, and weight, were collected by regular interviews. The data of the INR and warfarin doses were collected by a pharmacist during the treatment in hospital and in the anticoagulation clinic for discharged patients during the follow up.

553

2.4. Genotyping Polymorphisms of CYP2C9*3 (1075G N T, rs1057910) and VKORC1 (1639G N A, rs9923231) were detected by using DNA microarrays. Whole blood obtained from the forearm vein was collected after the patients registered for the study before surgery. Genomic DNA was extracted by a DNA Blood mini kit (Baio Technology Co, Ltd, Shanghai, China), as described by the product. Gene sequences were amplified by PCR. Mutant alleles were detected by using genotyping microarray (Baio Technology Co, Ltd, Shanghai, China). 2.5. Follow-up A one more year long follow-up was accomplished by the Anticoagulation Clinic of the hospital at least once every 3 months after the patients were discharged, and the INRs of the patients were monitored during the phase of follow-up. The dosage of warfarin was adjusted to keep the therapeutic INR within the range of 1.8-2.5. During the follow-up period, medication, dietary habits, and other information were also collected. 2.6. Selection and Comparison of Algorithms We searched for articles reporting pharmacogenetics algorithms for warfarin dosing through PubMed. MeSH terms were used as follows: ‘warfarin,’ ‘pharmacogenetics,’ ‘CYP2C9,’ and ‘VKORC1.’ The selected algorithms needed to include clinical and genetic (CYP2C9 and VKORC1 only) variables. We selected algorithms derived from mixed race, Caucasians, Asians (especially East Asian) and Han-Chinese to compare the influence of population to the performance. We chose the algorithm of Kim et al., which incorporate chronic heart function and derived from patients undertaking heart valve replacement, to evaluate contribution of heart function. Since we only tested SNPs of CYP2C9 and VKORC1, the algorithms with genetic variables other than CYP2C9 and VKORC1 were excluded. Algorithms with a sample size less than 100 patients were also excluded. To compare the performance of the algorithms, the MAE (mean absolute error) and percentage of patients whose predicted warfarin dose fell within 20% of the actual therapeutic dose in both initial and stable phases were considered in accordance to other studies [10]. The MAE was defined as the mean absolute difference between the predictive dose and actual therapeutic dose. The percentage of patients whose predictive warfarin dose was below (underestimation), above (overestimation) and within (ideal dose) the 20% of actual therapeutic dose was calculated to evaluate the accuracy of the algorithms. The predicted warfarin doses of the patients were calculated by using the algorithms established by each study and the collected data of the patients. The genotype of VKORC1 1639G N A was used if the algorithm contained VKORC1 polymorphism other than VKORC1 1639G N A, but in a strong linkage disequilibrium with it [26]. The performance of the algorithms was also evaluated in subgroups: (1) a low dose group (1.88 mg/day), an ideal dose group (1.88–4.38 mg/day), and a high dose group (4.38 mg/day) [15,27,28], (2) Patients with chronic heart failure (CHF) and without CHF, and (3) patients older than 50 (≥ 50) and younger than 50 (b 50). 2.7. Statistical Analysis The frequency distribution, percentage distribution, mean, and standard deviation were calculated as descriptive statistics. The Hardy-Weinberg equilibrium was used to evaluate the distribution of each genotype by using a chi-square test. Comparison between two groups was conducted by a t-test for continuous variables. Analysis of the statistics was completed by the Statistical Package for Social Science (SPSS ver. 18.0, SPSS Science, Chicago, IL, USA). A p-value of b0.5 was considered statistically significant.

554

H. Xu et al. / Thrombosis Research 136 (2015) 552–559

3. Results

Table 1 Main characters of the study population.

In total, 212 patients with written informed consent undertaking mechanical heart valve replacement were registered in the study from July 2012 to December 2012. A cohort of 193 patients (Fig. 1) with an average age of 53.16 ± 12.02 remained, including 92 males (47.42%) and 101 females (52.58%), while the others were excluded. Among the excluded patients, 5 patients gave up the surgery, 9 patients had abnormal liver function or renal impairment in the postoperative period, 3 patients transferred to another hospital, and 2 patients died after the surgery. A summary of the characters of the patients is shown in Table 1. The relation between warfarin dose and each genotype is summarized in Table 2. The dose of warfarin, in the stage of initial therapy and stable anticoagulation, respectively, shows significant differences across the genotypes of CYP2C9 (Initial, P = 0.002 and stable, P = 0.003) and VKORC1 (Initial, P b 0.001 and stable, P = 0.028). A one-year follow-up was conducted for patients (103 patients) with mechanic heart valve replacement (Fig. 1). 18 patients who had poor compliance of prescription of warfarin or INR test were excluded. And, we lost contact with 26 patients. Finally, the data of 59 patients was collected for the validation of the stable period. 3.1. Selection of Algorithms In total, 10 algorithms were selected for the validation of the performance. The IWPC [10] and Gage et al. algorithms [11] were derived from mixed races. The Wadelius et al. algorithm [12] was based on the Caucasian population, and 7 algorithms [13–19] based on Asian population were included for the study.

Variables

Value(n = 193)

Male, n(%) Age, years, mean ± SD Height, cm, mean ± SD Weight, kg, mean ± SD EF (%), mean ± SD Comorbidities, n (%) PVT Stroke Hypertension Heart failure Smoking, n (%) Surgery, n (%) Mechanic valve replacement AVR MVR DVR Bioprosthetic valve replacement Heart valvuloplasty Concomitant medications, n (%) ACEI/ARB β-blocker Calcium channel blockers Statins Amiodarone Digoxin Diuretics CYP2C9*3, n(%) *1/*1 *1/*3 VKORC1, n(%) AA AG GG

92 (47.67%) 53.16 ± 12.02 163.39 ± 8.69 61.58 ± 11.42 52.93 ± 5.15 17 (8.81%) 15 (7.77%) 40 (20.73%) 94 (48.70%) 32 (16.58%)

30 (15.54%) 39 (2/.21%) 34 (17.62%) 57 (29.53%) 33 (17.10%) 74 (38.34%) 53 (27.32%) 15 (7.77%) 16 (8.29%) 28 (14.51%) 61 (31.61%) 189 (97.93%) 184 (95.34%) 9 (4.66%) 157 (81.35%) 34 (17.62%) 2 (1.03%)

PVT, prior deep venous thrombosis or pulmonary embolism; AVR, aortic valve replacement; MVR, mitral valve replacement; DVR, double valve replacement.

3.2. Validation of Selected Algorithms Overall, the dose predicted by all selected pharmacogenetics algorithms fell within the 20% threshold of therapeutic dosage with the average percentage of 44.0 ± 8.8% (35.2%-52.8%) and 44.6 ± 9.7% (34.9%-54.3%) for the initial and stable phases, respectively. The MAE of the total selected algorithms was 0.85 ± 0.18 mg/day (0.67-1.03 mg/day) and 0.93 ± 0.19 mg/day (0.74-1.12 mg/day) for the initial and stable phases, respectively. In the phase of initial anticoagulation, the algorithms established by Huang et al. [19], Tan et al. [15], and Gage et al. [11] had better performance with the percentage within

20% of 50.8%, 49.7% and 49.7% and MAEs of 0.74 mg/day, 0.79 mg/day, and 0.76 mg/day, respectively (Table 4). In the phase of stable anticoagulation, the algorithms of IWPC [10], Ohno et al. [14], and Gage et al. [11] had higher accuracy of warfarin dose prediction with the percentage within 20% of 55.9%, 52.5%, and 50.8% and MAEs of 0.77 mg/day, 0.73 mg/day, and 0.76 mg/day, respectively (Table 5). However, the performance of Wadelius et al. [12] and Miao et al. [17] was poor with percentages within 20% of 29.9% and 25.9% in the initial

212 patients involved

193 patients for Initial validation

90 patients without mechanical valve

19 patients excluded

103 patients with mechanical valves

59 patients for stable validation

44 patients excluded

Fig. 1. Patients for validation.

H. Xu et al. / Thrombosis Research 136 (2015) 552–559 Table 2 Dose for each genotype (mean ± SD, mg/day). Genotype CYP2C9 ⁎1/⁎1 ⁎1/⁎3 VKORC1 AA AG GG

Initial dose

555

Table 4 Performance of selected algorithms in initial phase. Maintenance dose

3.11 ± 1.22 1.83 ± 0.69

3.50 ± 1.23 2.61 ± 0.50

2.75 ± 0.97 4.22 ± 1.28 6.75

3.23 ± 1.04 4.17 ± 1.10 6.75

P value

Algorithm

Underestimation

ideal

overestimation

MAE

0.741⁎ 0.002⁎⁎ 0.003⁎⁎⁎ 0.910⁎ b0.001⁎⁎ 0.028⁎⁎⁎

IWPC Gage et al. Wadelius et al. Kim et al. Ohno et al. Tan et al. You et al. Miao et al. Wen et al. Huang et al.

12.4 21.2 2.6 21.2 10.9 33.2 27.5 65.8 36.3 25.4

47.7 49.7 29.9 42.5 48.7 49.7 47.7 25.9 47.2 50.8

39.9 29.1 77.5 26.3 40.4 17.1 24.8 8.3 16.5 23.8

0.81 ± 0.65 0.76 ± 0.64 1.28 ± 0.85 0.82 ± 0.65 0.76 ± 0.58 0.79 ± 0.71 0.72 ± 0.58 1.05 ± 0.79 0.79 ± 0.70 0.74 ± 0.62

⁎ P-values were the results for χ2 test of evaluating Hardy-Weinberg equilibrium. ⁎⁎ Represents the comparison between the different polymorphisms of CYP2C9 or VKORC1 in the initial phase. ⁎⁎⁎ Represents the comparison between the different polymorphisms of CYP2C9 or VKORC1 in the stable phase.

treatment and 28.8% and 27.1% in stable treatment. In the phase of initial therapy, 5 algorithms had a larger percentage of overestimation than that of underestimation, while 4 algorithms had a larger percentage of overestimation than underestimation in the phase of stable treatment. Algorithms derived from Han-population had larger part of underestimation than overestimation in both initial and stable phase. (See Table 3.) For further subgroup studies, patients were stratified by different dose ranges (Dose b 1.88 mg/day, 1.88 b dose b 4.38 mg/day and Dose b 4.38 mg/day). All algorithms but the Miao et al. [17] and the Wadelius et al. algorithms [12] had higher predictive accuracy in the intermediate dose group in both periods of anticoagulation therapy. However, the algorithms of Wadelius et al. [12] (61.3% and 53.3%) and Miao et al. [17] (35.5% and 50.0%) had better performance in the high dose group and low dose group, respectively, in both periods of warfarin therapy. The Wen et al. [18] and Huang et al. algorithms [19] had a higher predictive accuracy of 64.1% in the initial phase, while IWPC [10] and Ohno et al. [14] performed better in the stable phase with the predictive percentage of 61.9% in intermediate dose group. The Miao et al. [17] algorithm had better performance in the low dose group with the percentage within 20% of 35.5% and 50.0% for the initial and stable phase, respectively. Details are summarized in Fig. 2. Patients were also stratified by heart function for further study of subgroups. Average warfarin doses were 3.11 ± 1.32 mg/day and 2.99 ± 1.14 mg/d for the groups of CHF and normal heart function, separately. A summary of details is shown in Fig. 3. In the initial phase of warfarin treatment, the Gage et al. algorithm [11] had better performance with 51.1% of patients with appropriate predicted dose in the group of CHF. The Huang et al. algorithm [19] had more accuracy in the group of patients with normal heart function with 53.5% within 20% of the therapeutic dose. However, Miao et al. [17] and Wadelius [12] had poor performance in both two groups. The selected algorithm had better performance in the group of normal function with the percentage of 45.2 ± 9.7%, and the percentage of patients in the group of CHF was 43.1% ± 9.4%. However, the risk of overestimation of patients with normal heart function was higher with a percentage of

34.5% ± 19.0% within 20% of the actual dose compared with that of patients with CHF, which was 26.8 ± 14.9%. The Kim et al. algorithm [13], IWPC [10], and Gage et al. algorithm [11] had better performance in the CHF group in both periods. However, other algorithms derived from the Asian population had better performance in the Normal group in the initial period. (See Fig. 4.) For patients in the phase of stable warfarin therapy, the average doses were 3.50 ± 1.35 mg/day for the group of CHF and 3.37 ± 0.93 mg/d for the group with normal heart function, separately. The IWPC (58.1%) [10] performed better in the group of patients with CHF. The Wen et al. algorithm [18] (60.7%) took advantage in the group of normal heart function. Patients in the group of CHF had higher risk for underestimation with 38.1% ± 19.5% within actual dose when compared with that of 23.7% ± 19.4% for the other group. Higher risk of overestimation can be observed in the group of normal heart function (30.3% ± 22.0% vs. 18.1 ± 15.1%). Patients were divided into two groups by age. In the phase of initial anticoagulation, Tan et al. [15] and Huang et al. [19] had more accuracy with the 55.7% and 56.9% within 20% of the therapeutic dose, respectively. All selected algorithms had better performance in the group of elderly patients for the period of stable anticoagulation. More than 60% of the patients’ actual doses were accurately predicted by IWPC [10], Gage et al. [11], Ohno et al. [14] and Tan et al. algorithms [15]. 4. Discussion In this study, performance of genotype-based warfarin dosing algorithms was validated in a cohort of 193 patients undertaking surgery of heart valve replacement and heart valvuloplasty in the periods of initial and stable treatment. In total, 10 genotype-based warfarin dosing algorithms were selected. Five algorithms were derived from the Han population, one algorithm was based on Japanese, and one was based on the Korean population. The algorithms of IWPC [10] and Gage et al. [11] were derived from diverse ethnicities, and Wadelius et al. [12] was derived from Gaussians, based on large sample sizes. All algorithms but the Wadelius et al. [12] and Miao et al. algorithms [17] had similar performance in the initial and stable phases

Table 3 Algorithms selected for evaluation. Algorithm

Ethnicity

IWPC [10]

Diverse

Gage et al. [11] Wadelius et al. [12] Kim et al. [13] Ohno et al. [14] Tan et al. [15] You et al. [16] Miao et al. [17] Wen et al. [18] Huang et al. [19]

Diverse Caucasians Korean Japanese Han Han Han Han Han

Target INR 2-3 1.5-2.8 2-3 1.7-2.8 1.5-3.0 1.7-3.0 2.0-3.5 1.5-3.0 1.7-3.0 1.8-3.0

Clinical variables

R2 (%)

Age,height,weight,VKORC1,CYP2C, race,enzyme inducer, amiodarone VKORC1,CYP2C9,BSA,age,INR,amiodarone,smoking,PVT VKORC1,CYP2C9,age,gender, number of INR-increasing drugs VKORC1,CYP2C9,age,weight,CHF, INR-increasing drug,INR-decreasing diet VKORC1,CYP2C9,BSA,age VKORC1,CYP2C9,BSA,age,number of INR-increasing drugs,smoking, stroke, hypertension VKORC1,CYP2C9,age,weight, number of Vk food intake age, weight, VKORC1, CYP2C9 VKORC1, CYP2C9, age, BSA age, BSA, VKORC1,CYP2C9

43 54 58.7 56 54.8 56.4 68 62.8 62 54.1

556

H. Xu et al. / Thrombosis Research 136 (2015) 552–559

Table 5 Performance of selected algorithms in stable phase. Algorithm

Underestimation

ideal

overestimation

MAE

IWPC Gage et al. Wadelius et al. Kim et al. Ohno et al. Tan et al. You et al. Miao et al. Wen et al. Huang et al.

15.3 25.4 5.1 16.9 20.3 32.2 45.8 71.2 45.8 40.1

55.9 50.8 28.8 45.8 52.5 45.8 40.7 27.1 49.2 49.2

28.8 23.8 66.1 37.3 27.2 22.0 13.5 1.7 5.0 10.7

0.77 ± 0.67 0.77 ± 0.66 1.18 ± 0.80 1.04 ± 0.81 0.73 ± 0.69 0.82 ± 0.71 0.89 ± 0.78 1.30 ± 0.98 0.94 ± 0.83 0.87 ± 0.76

of anticoagulation treatment. The Gage et al. algorithm [11] had better performance in both phases of initial and stable treatment. Compared with algorithms derived from Asian and mixed populations, the Wadelius et al. algorithm [12] had poor performance. The Warfarin maintenance dose (3.45 ± 1.18 mg/day) tended to increase compared with the initial warfarin dose (3.05 ± 1.23 mg/day). Sensitivity to warfarin would increase in early stage of post-surgery [24]. It may be caused by changes of hepatic or renal function, hypovolemia, serum albumin concentration, and habitat of daily diet. However, the 10 selected algorithms had similar performance (44.0 ± 8.8% vs. 44.6 ± 9.7%) in the phase of initial and stable treatment in total. In the initial phase of warfarin treatment, Huang et al. [19] had the best performance with the highest accuracy of the prediction of warfarin dose within 20% of the actual therapeutic dose. In the stable phase of warfarin treatment, the IWPC [10] algorithm had the largest predictive proportion of warfarin dose that fell within the 20% threshold of actual warfarin maintenance dose. The Gage et al. algorithm [11] had higher predictive accuracy in both the initial and stable period. The Wadelius et al. [12] (29.9% and 28.8%) and Miao et al. [17] (25.9% and 27.1%) algorithms had poor performance in both periods. The unsatisfactory

accuracy of these two algorithms (Miao 1.5-3.0 and Wadelius 2.0-3.0) might due to the wider range of target INR compared to our study. Similar results for poor performance of the Miao et al. algorithm [17] were obtained in other articles [15,29]. In contrast, the Miao et al. algorithm [17] performed well in the research of Zhao [28]. The performance of the algorithms differed by the different populations they derived from. Algorithms derived from Chinese, Japanese, and Korean populations had similar predictive accuracy on the warfarin dose except for the Miao et al. algorithm [17]. Algorithms (IWPC [10] and Gage [11]) of diverse populations and large sample sizes also had similar performance on the warfarin dose prediction and performed even better in the stable warfarin treatment phase. Liu [29] also obtained the conclusion that algorithms derived from Asian population performed no less than algorithms derived from mixed populations. However, Zhao [28] found that algorithms derived from Han population had better performance than algorithms derived from mixed populations. Shin [20] had the result that mixed population algorithms had better performance than populations that were not mixed. Consistent with our study, algorithms derived from Caucasians had poor performance [29]. The main cause of the poor performance of the Miao et al. algorithm [17] may be that the Target INR was 1.5-3.0. The Miao et al. algorithm [17] still had poor performance in validation of Liu’s [29] study with a target INR of 1.6-2.5 but performed well in Zhao’s validation [28] with a Target INR of 1.5-2.5. For Asian population, risk of bleeding is higher than that of Caucasian population [31]. It is indicated [32] that low intensity anticoagulation is appropriate for Chinese which differs from the target INR of Caucasians [2]. The commonly relatively ideal accuracy of algorithms derived from Asian population might be the similar target range of INR of derivation cohort. And the essential difference between races might also contribute to the performance diversity. The findings above indicate that the algorithms derived from specific popoulation are more preferred for local use, especially for Asians who need low intensity anticoagulation therapy with higher risk of bleeding. Meanwhile, algorithms derived from mixed race based on large sample

Fig. 2. Performance of algorithms in different warfarin dose range in initial period (A) and stable period (B). Low: Dose ≤ 1.88 mg/day; Intermediate: 1.88 b dose ≤ 4.38 mg/day; High: Dose N 4.38 mg/day.

H. Xu et al. / Thrombosis Research 136 (2015) 552–559

557

Fig. 3. Performance of algorithms in patients with different heart function in initial period (A) and stable period (B).

size also have high value for clinical application. And it may be not appropriate to select algorithms derived from other specific population. In the subgroup analysis, the main reasons for low accuracy were the high proportion of overestimation for the Wadelius et al. algorithm [12] and the high proportion of underestimation for the Miao et al. algorithm [17]. However, the Wadelius et al. algorithm [12] had

higher accuracy in the high dose group (Dose N 4.38 mg/day) while the Miao et al. algorithm [17] had higher accuracy in the low dose group (Dose b 1.88 mg/day). Other algorithms had high predictive accuracy (N50%) in the intermediate dose group (1.88 b dose b 4.38 mg/day). The Huang et al. algorithm [19] (64.1%) had the best performance in initial period, while the IWPC [10] and Ohno et al. algorithms [14] had an

Fig. 4. Performance of algorithms in different age groups in initial period (A) and stable (B) period.

558

H. Xu et al. / Thrombosis Research 136 (2015) 552–559

accurate prediction proportion of 61.9% in the stable period. Algorithms derived from diverse population also had better performance in the high dose group. The research of Liu and Zhao [28] obtained similar results. Algorithms derived from the Asian population mainly performed well in the intermediate dose group. Similar results were found in Zhao’s [28] research. Patients undertaking heart valve surgery have a high prevalence of CHF. The performance of algorithms in patients with different heart function remained to be validated. Algorithms derived from diverse populations performed better in the group of patients with CHF in both phases of anticoagulation therapy. In the stable treatment period, selected algorithms had a higher risk of underestimation in the group of CHF and overestimation in the group with normal heart function. Kim et al. [13] incorporate CHF into their algorithm as a variable. The Kim et al. algorithm [13] had better performance in the group of patients with CHF (51.6%) than in the group of patients with normal heart function(39.3%)in the phase of stable anticoagulation therapy. Patients with CHF decrease the hepatic hemoperfusion, which can increase the anticoagulation effect of warfarin. Sensitivity of patients to warfarin will increase at early stage of post surgery. However, the combination of medication could decrease the predicted dose of algorithms if the variables of combination drugs are included. Age is an important variable in all selected algorithms. Algorithms derived from patients of different age ranges may have different performance. For further study, patients were divided into two groups by age. The selected algorithms had better performance in the group of patients older than 50 (51.0% ± 15.4%) than the other group (34.2% ± 6.7%) in the stable phase of anticoagulation. The algorithms had relatively poor performance in the group of patients age b 50. The You et al. algorithm [16] was derived from patients with age N50. The algorithm had better performance in the elderly group during the stable phase. The You et al. algorithms [16] also had poor performance in patients b50 in the stable phase of treatment. Patients of Ohno et al. [14] had an average age of more than 70 years, and patients of the IWPC [10] and Gage et al. [11] algorithms had a relatively high average age. The algorithms had better performance in both periods. Overall, the differences of performance among the selected algorithms were mainly caused by several reasons as follows: derived from different ethnicities; consisted of patients with different indications; differences in target INR for anticoagulation treatment; differences in variables of algorithms; differences in form of the same variable in the algorithm (for example, drug interactions and number of drugs that increase INR serve as a variable in the Tan algorism [15], while this variable changes to whether the patient had a drug that increased INR in the Kim algorithm [13] and detail drug (amiodaron) in IWPC algorithm [10]). It is more reasonable to set detail drugs as variables to improve the accuracy of the algorithm. Pharmacogenetics-based algorithms can reduce the time to reach therapeutic INR and increase the time in therapeutic range [30]. Algorithms can be improved and new algorithms with higher prediction accuracy can be accomplished by validation of the published algorithms. However, our study had some limitations as follows. The study was a single-centre research with only 59 patients for algorithms performance validation in stable phase, which lead to relatively insufficient sample size especially for low and high dose group, a multi-centre study with large sample size to validate the performance of the algorithms and ascertain our findings. Furthermore, the study only tested SNPs of CYP2C9 and VKORC1, so we chose the algorithms with variables of CYP2C9 and VKORC1 polymorphisms. And the algorithms consist of the other genetic variants, like CYP4F2 which was not tested, could not be validated in the present study. Algorithms that incorporate CYP4F2 had better performance in the studies of Liu [29] and Zhao [28]. So further studies that compare the performance between algorithms incorporate different SNPs are needed. Thirdly, the study validated selected algorithms only in patients with heart valve surgery, and more studies which compare the performance of algorithms in patients

with other specific diseases. At last, it is important to note that further study focus on the outcome of genetic-based algorithms need to be conducted for clinical application. The algorithms are anticipated and established by incorporating detailed data of patients with heart valve surgery to improve the anticoagulation treatment. Our research found that nearly all selected algorithms had similar performance, and no algorithms had the best performance in all subgroups. It indicated that special clinical characters of patients undertaking heart valve surgery should be considered to select appropriate algorithm. Alternatively, new algorithms can be published for such kinds of patients to incorporate specific variables. Disclosures The authors have no any conflict of interest to disclose. Acknowledgments The author would like to thank all patients for their efforts in this research. The author also would like to thank the support of Department of Cardiothoracic Surgery, Affiliated Drum Tower Hospital of Nanjing University Medical School. References [1] A. Tomek, V. Maťoška, T. Kolářová, et al., The bleeding risk during warfarin therapy is associated with the number of variant alleles of CYP2C9 and VKORC1 genes, Cardiology 125 (2013) 182–191. [2] P. Monagle, A.K. Chan, N.A. Goldenberg, et al., Antithrombotic therapy in neonates and children: Antithrombotic Therapy and Prevention of Thrombosis, 9th ed: American College of Chest Physicians Evidence-Based Clinical Practice Guidelines, Chest 141 (2012) e737S–e801S. [3] N. Allou, P. Piednoir, C. Berroëta, et al., Incidence and risk factors of early thromboembolic events after mechanical heart valve replacement in patients treated with intravenous unfractionated heparin, Hear 95 (2009) 1694–1700. [4] W. Zhang, W.J. Zhang, J. Zhu, et al., Genetic polymorphisms are associated with variations in warfarin maintenance dose in Han Chinese patients with venous thromboembolism, Pharmacogenomics 13 (2012) 309–321. [5] T.S. Perlstein, S.Z. Goldhaber, K. Nelson, et al., The Creating an Optimal Warfarin Nomogram (CROWN) Study, Thromb. Haemost. 107 (2012) 59–68. [6] G. D'Andrea, R.L. D'Ambrosio, P. Di Perna, et al., A polymorphism in the VKORC1 gene is associated with an interindividual variability in the dose-anticoagulant effect of warfarin, Blood 105 (2005) 645–649. [7] M. Wei, F. Ye, D. Xie, et al., A new algorithm to predict warfarin dose from polymorphisms of CYP4F2, CYP2C9 and VKORC1 and clinical variables: Derivation in Han Chinese patients with non valvular atrial fibrillation, Thromb. Haemost. 107 (2012) 1083–1091. [8] P. Lenzini, M. Wadelius, S. Kimmel, et al., Integration of genetic, clinical, and INR data to refine warfarin dosing, Clin. Pharmacol. Ther. 87 (2010) 572–578. [9] C.F. Zambon, V. Pengo, R. Padrini, et al., VKORC1, CYP2C9 and CYP4F2 genetic-based algorithm for warfarin dosing: an Italian retrospective study, Pharmacogenomics 12 (2011) 15–25. [10] International Warfarin Pharmacogenetics Consortium, T.E. Klein, R.B. Altman, et al., Estimation of the warfarin dose with clinical and pharmacogenetic data, N. Engl. J. Med. 360 (2009) 753–764. [11] B.F. Gage, C. Eby, J.A. Johnson, et al., Use of pharmacogenetic and clinical factors to predict the therapeutic dose of warfarin, Clin. Pharmacol. Ther. 84 (2008) 326–331. [12] M. Wadelius, L.Y. Chen, J.D. Lindh, et al., The largest prospective warfarin-treated cohort supports genetic forecasting, Blood 113 (2009) 784–792. [13] H.S. Kim, S.S. Lee, M. Oh, et al., Effect of CYP2C9 and VKORC1 genotypes on earlyphase and steady-state warfarin dosing in Korean patients with mechanical heart valve replacement, Pharmacogenet. Genomics 19 (2009) 103–112. [14] M. Ohno, A. Yamamoto, A. Ono, et al., Influence of clinical and genetic factors on warfarin dose requirements among Japanese patients, Eur. J. Clin. Pharmacol. 65 (2009) 1097–1103. [15] S.L. Tan, Z. Li, G.B. Song, L.M. Liu, et al., Development and comparison of a new personalized warfarin stable dose prediction algorithm in Chinese patients undergoing heart valve replacement, Pharmazie 67 (2012) 930–937. [16] J.H. You, R.S. Wong, M.M. Waye, et al., Warfarin dosing algorithm using clinical, demographic and pharmacogenetic data from Chinese patients, J. Thromb. Thrombolysis 31 (2011) 113–118. [17] L. Miao, J. Yang, C. Huang, et al., Contribution of age, body weight, and CYP2C9 and VKORC1 genotype to the anticoagulant response to warfarin: proposal for a new dosing regimen in Chinese patients, Eur. J. Clin. Pharmacol. 63 (2007) 1135–1141. [18] M.S. Wen, M. Lee, J.J. Chen, et al., Prospective study of warfarin dosage requirements based on CYP2C9 and VKORC1 genotypes, Clin. Pharmacol. Ther. 84 (2008) 83–89. [19] S.W. Huang, H.S. Chen, X.Q. Wang, et al., Validation of VKORC1 and CYP2C9 genotypes on interindividual warfarin maintenance dose: a prospective study in Chinese patients, Pharmacogenet. Genomics 19 (2009) 226–234.

H. Xu et al. / Thrombosis Research 136 (2015) 552–559 [20] J. Shin, D. Cao, Comparison of warfarin pharmacogenetic dosing algorithms in a racially diverse large cohort, Pharmacogenomics 12 (2011) 125–134. [21] P.B. Shaw, J.L. Donovan, M.T. Tran, et al., Accuracy assessment of pharmacogenetically predictive warfarin dosing algorithms in patients of an academic medical center anticoagulation clinic, J. Thromb. Thrombolysis 30 (2010) 220–225. [22] J.P. Rose, T.L. Rihn, S.F. Long, Warfarin sensitivity after mechanical heart valve replacement, Pharmacotherapy 18 (1998) 856–859. [23] W. Ageno, A.G. Turpie, Exaggerated initial response to warfarin following heart valve replacement, Am. J. Cardiol. 84 (1999) 905–908. [24] M. Rahman, T.M. BinEsmael, N. Payne, et al., Increased sensitivity to warfarin after heart valve replacement, Ann. Pharmacother. 40 (2006) 397–401. [25] K1. Meijer, Y.K. Kim, S. Schulman, Decreasing warfarin sensitivity during the first three months after heart valve surgery: implications for dosing, Thromb. Res. 125 (2010) 224–229. [26] N.A. Limdi, M. Wadelius, L. Cavallari, et al., Warfarin pharmacogenetics: a single VKORC1 polymorphism is predictive of dose across 3 racial groups, Blood 115 (2010) 3827–3834.

559

[27] Q. Xu, B. Xu, Y. Zhang, et al., Estimation of the warfarin dose with a pharmacogenetic refinement algorithm in Chinese patients mainly under low-intensity warfarin anticoagulation, Thromb. Haemost. 108 (2012) 1132–1140. [28] L. Zhao, C. Chen, B. Li, et al., Verification of pharmacogenetics-based warfarin dosing algorithms in Han-Chinese patients undertaking mechanic heart valve replacement, PLoS One 9 (2014) e94573. [29] Y. Liu, J. Yang, Q. Xu, et al., Comparative performance of warfarin pharmacogenetic algorithms in Chinese patients, Thromb. Res. 130 (2012) 435–440. [30] M. Pirmohamed, G. Burnside, N. Eriksson, et al., A randomized trial of genotypeguided dosing of warfarin, N. Engl. J. Med. 369 (2013) 2294–2303. [31] S. Suzuki, T. Yamashita, T. Kato, et al., Incidence of major bleeding complication of warfarin therapy in Japanese patients with atrial fibrillation, Circ. J. 71 (2007) 761–765. [32] J.H. You, F.W. Chan, R.S. Wong, et al., Is INR between 2.0 and 3.0 the optimal level for Chinese patients on warfarin therapy for moderate-intensity anticoagulation? Br. J. Clin. Pharmacol. 59 (2005) 582–587.