- Research article
- Open Access
- Open Peer Review
A prediction study of warfarin individual stable dose after mechanical heart valve replacement: adaptive neural-fuzzy inference system prediction
© The Author(s). 2018
- Received: 8 September 2017
- Accepted: 1 February 2018
- Published: 15 February 2018
It’s difficult but urgent to achieve the individualized rational medication of the warfarin, we aim to predict the individualized warfarin stable dose though the artificial intelligent Adaptive neural-fuzzy inference system (ANFIS).
Our retrospective analysis based on a clinical database, involving 21,863 patients from 15 Chinese provinces who receive oral warfarin after the heart valve replacement. They were allocated into four groups: the external validation group (A group), the internal validation group (B group), training group (C group) and stratified training group (D group). We used a univariate analysis of general linear models(GLM-univariate) to select the input variables and construct two prediction models by the ANFIS with the training and stratified training group, and then verify models with two validation groups by the mean squared error(MSE), mean absolute error(MAE) and the ideal predicted percentage.
A total of 13,639 eligible patients were selected, including 1639 in A group, 3000 in B group, 9000 in C group, and 3192 in D group. Nine input variables were selected out and two five-layered ANFIS models were built. ANFIS model achieved the highest total ideal predicted percentage 63.7%. In the dose subgroups, all the models performed best in the intermediate-dose group with the ideal predicted percentage 82.4~ 86.4%, and the use of the stratified training group slightly increased the prediction accuracy in low-dose group by 8.8 and 5.2%, respectively.
As a preliminary attempt, ANFIS model predicted the warfarin stable dose properly after heart valve surgery among Chinese, and also proved that Chinese need lower anticoagulation intensity INR (1.5–2.5) to warfarin by reference to the recommended INR (2.5–3.5) in the developed countries.
- Rational medication
- Heart valve surgery
- Adaptive neural-fuzzy inference system
- Dose prediction
Warfarin is the most commonly prescribed anticoagulant with its best effectiveness and low-cost for long-term management with cardiac valve replacement [1, 2]. However, the use of warfarin is restricted by the narrow therapeutic window and highly individual variation, for instance, bleeding and embolism complications, resulting from the insufficient dose or overdose anticoagulation, led warfarin to be the main causes of severe adverse event worldwide [3, 4]. This fact is more apparent in Chinese by reference to the recommended INR(2.5–3.5) for developed countries. Therefore, the individual rational medication of warfarin dosage is the key to guarantee the safety and therapeutic effectiveness during the anticoagulation therapy.
Currently, numerous clinical and genetic prediction models have been applied to predict the warfarin stable dose in the anticoagulation therapy, most of which are based on multiple linear regression (MLR) [5–9]. To our acknowledgement, MLR has some well-known limitations in identifying and dealing with the nonlinear relationship between various influencing factors, which may cut down the prediction accuracy of the models. Hence, MLR might not be the most appropriate method to predict warfarin stable dose accurately, especially for the individuals. Adaptive Neuro-Fuzzy Inference System (ANFIS), as a new artificial intelligence, has strong skills of classification identifying complex nonlinear relationship and other possible interactions among predictors in the form of fuzzy rules quickly, and dealing with the nonlinear relationships between input variables by adaptive learning features. Previous studies have proved that ANFIS has been frequently applied in medical field successfully through establishing prediction models based on clinical database [10–15].
Hence, our study aims to predict the individual warfarin stable dose by ANFIS model based on a big clinical database retrospectively, which was generated by a registered and multicenter cohort with 21,863 cases have received the heart valve replacement in China and validate it with the internal and external validation groups.
Study design and setting of the study
We conducted our study retrospectively based on a database named “Chinese Low Intensity Anticoagulant Therapy after Heart Valve Replacement” (CLIATHVR), which was registry study involving 35 participating centers from 15 provinces in China collecting all patients who underwent heart valve replacement received anticoagulation with warfarin. We finally enrolled 21,863 collected cases from April 1, 2011 to November 5,2014, and authors had access to information that could identify individual participants during or after data collection after application. Written informed consent was obtained from all participants and the study was approved by the Ethics Committee of the West China hospital in Sichuan university with the number 2017(92).
Characteristics of participants
The inclusion criteria were as follows: age above 18 years, receive oral anti-coagulation warfarin in regular and monitor INR as index after receiving cardiac valve surgery; assure the fluctuation of INR less than 0.2 units for three times continuously during the later follow-up.
The exclusion criteria: Severe liver or kidney dysfunction before or after the operation; combined use of non-steroidal anti-inflammatory drugs or other drugs affecting coagulation; anticoagulant.
complications occurred during anticoagulant therapy (thrombosis, embolism, bleeding, death result).
Input variables and output variables
We identified 45 potential input variables among 706 items derived from the CLIATHVR database based on clinical experience and knowledge related to warfarin, including 16 items of demographic data; 17 items of preoperative medical examination data; 9 items of surgery type; 3 items of postoperative anticoagulation data. Due to the stability of screening the variables by the GLM-univariate analysis and the value of partial squared eta could reflect the influence of each variable, we adopted it to get the influential input variables from potential input variables. On the premise of presenting the prediction function of ANFIS models, we choose partial squared eta ≥0.002 in order to simplify the predicted models. The warfarin stable dose was set as the output variable.
We constructed four groups by eligible participants: the external validation group(A group), the internal validation group(B group), the training group(C group) and the stratified training group(D group). Since all the cases were arranged by the enrollment time, we extracted the last 12% enrolled cases of the whole sample as the external validation group (A Group). The rest of the eligible patients were randomly divided into the internal validation groups(B group)and the training (C group) and by the ratio of 1:3. Meanwhile, the training group (C group)was divided into three sub-groups based on the warfarin origin: the high-dose (warfarin stable dose ≥3.125 mg/d), intermediate-dose (1.875–3.125 mg/d) and low-dose (≤1.875 mg/d) group. Given the uneven distribution in the different dose subgroups, we created a stratified training group (D group) by using the method of equal random- stratified sampling, which randomly extracted the same number of patients from the other two dose groups equal to the smallest group. The training group (C group) and stratified training group (D group)were used to train models until achieve the appropriate parameters to construct the prediction model, and the external-(A group) and internal validation groups(B group) were used to verify the prediction accuracy of the model.
Modeling of ANFIS
Layer 1 is the fuzz-ification layer in which each node represents a membership value to a linguistic term as a Gaussian function with the mean;
Layer 2 provides the strength of the rule by means of multiplication operator. It performs AND operation;
Layer 3 is the normalization layer which normalizes the strength of all rules according to the equation;
Layer 4 is a layer of adaptive nodes. Every node in this layer computes a linear function where the function coefficients are adapted by using the error function of the multi-layer feed-forward neural network;
Layer 5 is the output layer whose function is the summation of the net outputs of the nodes in Layer 4.
Data measurement-model validation
Our primary outcome was the difference between predicted warfarin stable dose by ANFIS models and the actual warfarin stable dose in the clinical practice. The models’ predictive accuracy was evaluated by the external validation group(A group)and the internal validation(B group)with three indexes, including the mean squared error (MSE), mean absolute error (MAE) and the ideal predicted percentage. MAE is the average of absolute difference between the predicted dosage and the actual dosage that patients received. MSE is the square of the difference between two dosages. The ideal predicted percentage was defined as the percentage of patients whose predicted warfarin dose was within 20% of the actual dose.
Data was recorded by Microsoft Excel 2010. The comparisons of baseline characteristics of patients receiving warfarin between different groups when compared with the whole sample with the χ2 test for categorical variables and the independent sample t test for continuous variables. Difference between the predicted dose and the actual dose was analyzed by the paired t-test. All statistical tests were two-sided and a P value less than 0.05 was considered as statistically significant without exception. All the analyses were performed by SPSS 20.0.
Participants and descriptive data
the basic characteristic of the whole study
stratified training group
Left ventricular diastolic diameter
Inner diameter of leftatrium(mm)
Inner diameter of right atrium(mm)
Bleeding disease history
History of anticoagulant drugs
Mitral valve surgery
Tricuspid valve surgery
Aortic valve surgery
Left atrial appendage occlusion method
Left atrial volume loss
Non surgical bleeding
Origin of warfarin
Dosage form(2. 5 mg/tablet)
Starting time of anticoagulation(n days after surgery)
We finally determined nine variables derived from the 45 potential inputs by using the GLM-Univariate (partial η2 ≥ 0.002)method. These nine variables were identified, including age, NYHA classification, BSA (Body surface area), RAD (Right atrial diameter), Creatinine, APPT (Activated partial thromboplastin time),Radiofrequency ablation, Warfarin origin, Starting time of anticoagulation. The most influential variables in the screening were body surface area, warfarin origin and age. All the input variables were used to construct models and the warfarin stable dose was set as the outcome variable.
Comparison of predicted value and actual value of warfarin in the validation groups
Predicted value (mg/d)
Actual value (mg/d)
The ideal predicted percentage of warfarin in patients in the validation groups
The predictive ability of models in the different dose-subgroups
Low-dose group (N%)
Intermediate -dose group (N%)
High-dose group (N%)
The study showed that the ANFIS can predict the warfarin stable dose for Chinese patients after heart valve replacement surgery properly with the total ideal predicted percentage 57.0–63.7%. In the dose subgroups analysis, all the models performed best in the intermediate group with high ideal predicted percentage 80.1–86.4%. In the high dose group, ANFIS model achieved relatively high 25.1–40.2%, but the performance of models reduced in the stratified training, especially in the high dose group.
The International Warfarin Pharmacogenetics Consortium(IWPC)in 2009 established the most famous warfarin prediction model with 4043 eligible patients , the clinical model was built by clinical variables, and the genetic model was built by both the clinical variables and pharmacogenetic variables, and the 1009 patients comprised the validation cohort. Our prediction accuracy of the models in the internal validation is higher than IWPC clinical model except the lower predictive accuracy in low dose group (low dose group 0.3% vs 25.9%; intermediate dose group 86.4% vs 53.6%; high dose group 36.2% vs 9.6%; the total ideal predicted percentage 63.7% vs 39.3%).When considering the number distribution of different dose group, in IWPC model, low dose group didn’t occupy the smallest part, However, it hold the smallest in our model. Hence, the number distribution can affect the predicative accuracy especially the smallest subgroup. Besides, the MAE in our models is nearly less than 4.35 mg/week which is much lower than 9.9 mg/week compared to the clinical algorithm of the IWPC. The mean warfarin stable dose in our study nearly 20 mg/week was lower than the clinical model of IWPC with the dose 28 mg/week, which was consistent with the fact that Chinese need the lower INR during the anticoagulation therapy than westerns [17, 18]. Since the mean age of the patients in this study was 50 years and less than 1% of the patients had history of embolic disorders. Compared to other western world studies, the majority of the patients in our study did not have atherosclerosis as etiological factor, those obvious differences might explain the fact that Chinese population a lower dose of warfarin is needed in comparison to those recommended in the western world. Currently, the report of database in our multicenter registered study revealed that 88.6% of the target INR was between 1.5–2.5 and the mean INR was 1.84 ± 0.53 during hospitalization. All the above findings demonstrate the fact the low intensity with the INR range from 1.5 to 2.5 might be suitable for Chinese after the heart valve surgery.
How to identify the influential variables determines the stability and the prediction performance of the model, but whether the genetic variables can benefit patients or not is still controversial. We identified nine input variables by GLM-univariate analysis and variables is not completely consistent with the IWPC models. The adding of genetic variables in the IWPC pharmacogenetic model significantly improve the predictive accuracy of the subgroup dose, but the ideal predicted percentage in the intermediate group was 54.6%, which was lower than 86.4% in our study. A multicenter, randomized, controlled trial conducted in 2013  based on 455 patients concerning genotype-guided dosing of warfarin suggest that genotype-guided warfarin dosing was superior to standard dosing in the mean percentage of time(gene directs 67.4% vs the control group 60.3%), and the shorter median time needed to reach a therapeutic INR in the genotype-guided group. But another article published in the same year  with the same purpose indicates that the genotype-guided dosing of warfarin did not improve anticoagulation control during the first 4 weeks of therapy. Besides, the current genetic testing is expensive which isn’t in the scope of health insurance system service, and has not been popular in the developing countries. Hence, the above inconsistent results and issues would also arise if the genotype-guided dosing were implemented in the future research and clinical practice.
One important feature in our study is that we enrolled a large sample with 13,639 one single ethnic patients who received warfarin after heart valve replacement to establish and validate the models. Most previous study of warfarin prediction models based on different racial backgrounds, among which, the IWPC constructed prediction models with the maximum 4043 patients, including 55.2% White people, 30.4% Asian, 8.7% black and remaining mix population. Another important feature is the application of the internal and the external validation since that previous studies have proposed the internal validation may lead to optimism in prediction accuracy than external validation . Our design of the external validation group was a feasible attempt according to obvious statistical significance among some the basic characteristics and the warfarin dose compared to the internal validation group, and our results showed that the total prediction accuracy of external validation group is lower than the internal validation group by 3.1%. Based on our previous prediction modeling research conducted in 2014  and considering the linear relationship between independent variables and the dependent variable, we used the general linear models of univariate method to get the target input variables in the screening of variables.
We conducted this research rigorously, but there are still limitations to our study, one limitation was that we didn’t add the influential genetic information because of that we haven’t obtained it yet until now. In the selected input variables, the absence of major influencing factors in the difficult data collection, such as life style and drug combinations, could influence the prediction accuracy of the models by effecting the coagulation cascade. Thirdly, it is likely that some patients had more than one valve replaced, and the data we could obtained couldn’t meet our needs to do subgroup analysis. Lastly, using ANFIS, like any multidimensional analysis, importance is given to variables that have influence in that determined retrospective study. In other word, in this retrospective study, ANFIS is only a description of the phenomenon that has been analyzed, results should be validated in a prospective study in the future research before any clinical application.
In conclusion, our research explored the rational medication of individualized warfarin stable dose for Chinese patients after heart valve disease treatment adapting the ANFIS with patient-related clinical parameters based on 21,863 Chinese cases. The result indicated that ANFIS could predict the warfarin stable dose properly, which might play potential guiding role for doctors in deal with the Chinese patients, and also proved that Chinese need lower anticoagulation intensity INR (1.5–2.5) to warfarin by reference to the recommended INR (2.5–3.5) in the developed countries. Our database is large and being multicentric, reinforcement learning(RL) within machine learning might be tried in the further research with this ideal scenario, and genetic database will be combined with clinical variables to predict the individualized warfarin more accurately.
We greatly appreciate the support from other members of the CLIATHVR study team who provided their generous contribution of time and efforts help during the research.
The project was supported by National Natural Natural Science Foundation (81641021) and the National Science & Technology Pillar Program during the Twelfth Five-Year Plan Period (Project number 2011BAI11B18). The funding have no role in the design, implementation, data analysis, the article writing, and in the decision to submit the article for publication.
Availability of data and materials
HT and QL was first author and responsible for interpretation and discussion of data and drafting of the manuscript. QZ and JC responsible for performance of research, analysis of data and study methodology. BF, JLH and LD performed previous foundation work, data acquisition and clinical perspectives. WZQ and JW was responsible for performance of research and study quality control. JC and LD as the co-corresponding author, was responsible for the conception and design of research, revising critically for the important intellectual content and advised on data presentation, style and language and revised final manuscript. All authors agreed the version to be published.
Ethics approval and consent to participate
Written informed consent was obtained from all participants and the study was approved by the Ethics Committee of the West China hospital in Sichuan university with the number 2017(92).
Consent for publication
The authors declare that they have no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
- Ageno W, Gallus AS, Wittkowsky A, Crowther M, Hylek EM, Palareti G, American College of Chest P. Oral anticoagulant therapy: antithrombotic therapy and prevention of thrombosis, 9th ed: American college of chest physicians evidence-based clinical practice guidelines. Chest. 2012;141(2 Suppl):e44S–88S.View ArticlePubMedPubMed CentralGoogle Scholar
- Ansell J, Hirsh J, Hylek E, Jacobson A, Crowther M, Palareti G, American College of Chest P. Pharmacology and management of the vitamin K antagonists: American College of Chest Physicians Evidence-Based Clinical Practice Guidelines (8th edition). Chest. 2008;133(6 Suppl):160S–98S.View ArticlePubMedGoogle Scholar
- Wysowski DK, Nourjah P, Swartz L. Bleeding complications with warfarin use - a prevalent adverse effect resulting in regulatory action. Arch Intern Med. 2007;167(13):1414–9.View ArticlePubMedGoogle Scholar
- Budnitz DS, Lovegrove MC, Shehab N, Richards CL. Emergency hospitalizations for adverse drug events in older Americans. N Engl J Med. 2011;365(21):2002–12.View ArticlePubMedGoogle Scholar
- Grossi E, Podda GM, Pugliano M, Gabba S, Verri A, Carpani G, Buscema M, Casazza G, Cattaneo M. Prediction of optimal warfarin maintenance dose using advanced artificial neural networks. Pharmacogenomics. 2014;15(1):29–37.View ArticlePubMedGoogle Scholar
- Liang R, Li L, Li C, Gao Y, Liu W, Hu D, Sun Y. Impact of CYP2C9*3, VKORC1-1639, CYP4F2rs2108622 genetic polymorphism and clinical factors on warfarin maintenance dose in Han-Chinese patients. J Thromb Thrombolysis. 2012;34(1):120–5.View ArticlePubMedGoogle Scholar
- Pirmohamed M, Burnside G, Eriksson N, Jorgensen AL, Toh CH, Nicholson T, Kesteven P, Christersson C, Wahlstrom B, Stafberg C, et al. A randomized trial of genotype-guided dosing of warfarin. N Engl J Med. 2013;369(24):2294–303.View ArticlePubMedGoogle Scholar
- Li X, Liu R, Luo ZY, Yan H, Huang WH, Yin JY, Mao XY, Chen XP, Liu ZQ, Zhou HH, et al. Comparison of the predictive abilities of pharmacogenetics-based warfarin dosing algorithms using seven mathematical models in Chinese patients. Pharmacogenomics. 2015;16(6):583–90.View ArticlePubMedGoogle Scholar
- Liu R, Li X, Zhang W, Zhou HH. Comparison of nine statistical model based warfarin Pharmacogenetic dosing algorithms using the racially diverse international warfarin Pharmacogenetic consortium cohort database. PLoS One. 2015;10(8):e0135784.View ArticlePubMedPubMed CentralGoogle Scholar
- Avci E, Turkoglu I. An intelligent diagnosis system based on principle component analysis and ANFIS for the heart valve diseases. Expert Syst Appl. 2009;36(2):2873–8.View ArticleGoogle Scholar
- Buyukbingol E, Sisman A, Akyildiz M, Alparslan FN, Adejare A. Adaptive neuro-fuzzy inference system (ANFIS): a new approach to predictive modeling in QSAR applications: a study of neuro-fuzzy modeling of PCP-based NMDA receptor antagonists. Bioorg Med Chem. 2007;15(12):4265–82.View ArticlePubMedGoogle Scholar
- Polat K, Güneş S. An expert system approach based on principal component analysis and adaptive neuro-fuzzy inference system to diagnosis of diabetes disease. Digital Signal Processing. 2007;17(4):702–10.View ArticleGoogle Scholar
- Chang JJ, Syafiie S, Ahmad RKR, Lim TA. ANFIS Based Model for Bispectral Index Prediction. Recent Adv Soft Comput Data Mining. 2014;287:133–42.View ArticleGoogle Scholar
- Nilashi M, Ibrahim O, Ahani A. Accuracy improvement for predicting Parkinson's disease progression. Sci Rep. 2016;6:34181.View ArticlePubMedPubMed CentralGoogle Scholar
- Übeyli ED. Automatic diagnosis of diabetes using adaptive neuro-fuzzy inference systems. Expert Syst. 2010;27(4):259–66.View ArticleGoogle Scholar
- The International Warfarin Pharmacogenetics Consortium. Estimation of the warfarin dose with clinical and Pharmacogenetic data. N Engl J Med. 2009;360(8):753–64.View ArticlePubMed CentralGoogle Scholar
- Blann A, Hewitt J, Siddiqui F, Bareford D. Racial background is a determinant of average warfarin dose required to maintain the INR between 2.0 and 3.0. Br J Haematol. 1999;107:207–9.View ArticlePubMedGoogle Scholar
- Gan GG, Teh A, Goh KY, Chong HT, Pang KW. Racial background is a determinant factor in the maintenance dosage of warfarin. Int J Hematol. 2003;78:84–6.View ArticlePubMedGoogle Scholar
- Kimmel SE, French B, Kasner SE, Johnson JA, Anderson JL, Gage BF, Rosenberg YD, Eby CS, Madigan RA, McBane RB, et al. A pharmacogenetic versus a clinical algorithm for warfarin dosing. N Engl J Med. 2013;369(24):2283–93.View ArticlePubMedPubMed CentralGoogle Scholar
- Collins GS, de Groot JA, Dutton S, Omar O, Shanyinde M, Tajar A, Voysey M, Wharton R, Yu LM, Moons KG, et al. External validation of multivariable prediction models: a systematic review of methodological conduct and reporting. BMC Med Res Methodol. 2014;14:40.View ArticlePubMedPubMed CentralGoogle Scholar
- Zhou Q, Kwong J, Chen J, Qin W, Chen J, Dong L, Cmcs t. Use of artificial neural network to predict warfarin individualized dosage regime in Chinese patients receiving low-intensity anticoagulation after heart valve replacement. Int J Cardiol. 2014;176(3):1462–4.View ArticlePubMedGoogle Scholar