ISSUES

INTRODUCTION

Cardiovascular disease was recorded as the leading cause in 31% of all death, with more than 17.5 million deaths each year [1]. The number of cardiovascular disease including heart disease and stroke is expected to rise to more than 23.6 million by 2030 [2]. ​ The percentage of premature deaths (under the age of 70) from CVDs ranges from 4% in high-income countries to 42% in low-income countries [3]. Warfarin is the most widely used anticoagulant in the prevention and treatment of CVD. Warfarin is often used to prevent thrombosis and occlusion of blood flow disorders at fibrillation, Deep Venus Thrombosis (DVT), heart attacks, pulmonary embolism and heart valve replacement [4-7]. Due to large inter-individual variability in dose requirements, the warfarin dosage in anticoagulation is usually complicated and unpredictable [8, 9]. Under- or over- dosing will result in INR values outside the target range (2-3 for most conditions) and put the patient at risk of severe adverse outcomes such as bleeding or thromboembolism and it can exacerbate disease, hospitalization, and even death [4, 5, 10].

Many clinical and genetic features affect the dosage of patients [5, 9, 11] therefore, machine learning techniques can be used for the diagnosis or prediction of warfarin’s doses to help expert humans to make the best decisions [12, 13]. Artificial Neural Networks (ANNs) are one of the most important ML algorithms [14] which are apt for such applications as: first, they can incorporate large number of inputs and large number of outputs; second, they are dynamic and can be upgraded with new data sets periodically; third, they can explore the nonlinear relationship; forth, they have high generalizability [15, 16] and hence, can be used to predict warfarin dosage requirements.

In this paper, the Multilayer Perceptron (MLP) as common type of ANNs was used. Aim of this study is to find the best MLP structure based on conjugate gradient training algorithm. Then, the next step is to develop an intelligent system for prediction of optimal dose of warfarin that is suitable for Iranian adult patients with artificial heart valves.

DEVELOPMENT

First, a list of essential factors for warfarin dose determination was made by search the electronic databases and consultation with cardiologists. Then these factors were organized in the form of a questionnaire "Determining the importance of factors" and were given to cardiologists. Essential factors, whose method of extraction has already been published in [17] are: demographic information such as age, height and weight, and clinical information and lab test results include previous and current INR, target range of INR, previous and current dose of warfarin as output variable.

The data were extracted from the patient's medical records by using a random sampling method and was used to train and evaluate the MLP. The patients with artificial heart valves were treated in the second half of 2013 at the PT Tehran Heart Center were included.

In data preprocessing step, for each variable, missing values (31 data items) are completed by mean of that variable. Then, using the Larose normalization method, all the variables were normalized to the interval [0, 1]. In implementation step, several MLP neural networks with different number of neurons in hidden layer based on conjugate gradient training algorithm were constructed and evaluated to determine the best network for developing an intelligent system for warfarin dosing.

APPLICATION

Data of 846 patients including age, height, weight, previous and current INR, target range of INR, previous and current dose of warfarin as output variable were extracted. MLP neural networks with different structures were implemented in the MatLab programming environment. A total of nine MLP neural networks with different number of neurons in hidden layer (2 to 10) were created based on epochs=1000, goal =0, tangent hyperbolic activation function and conjugate gradient training algorithm. The performance of different neural networks was investigated using 10-fold cross validation and the results are compared based on disturbance rate, Mean Absolute Error (MAE) and regression coefficient (R2). The results revealed that, MLP with seven neurons in hidden layer is the best model with a MAE=0.1, disturbance rate=0.33 and R2=0.87. Therefore, this neural network was chosen to design the intelligent system for warfarin dosing. Graphical User Interface (GUI) of prediction system was created by using GUIDE in MatLab environment. Fig 1 shows an overview of data entrance page of the system.

[Figure ID: F1] Fig 1. Overview of data entrance page

DISCUSSION

There is a large body of literature describing the development of tools to predict warfarin dose requirements for adult patient in other countries around the world. The variables included in the prediction model, the implemented methods, and the findings of related studies are shown in Table 1.

CONCLUSION

According to previous studies, predicting warfarin dose based on clinical and demographic data only obtain performance as well as or even better than experts [18, 21]. However, Various studies have shown that adding genetic data improves the performance of prediction models [9, 19, 20], the proposed model, which predicts warfarin dose appropriately based on clinical data only, is preferred because collecting genetic data is time-consuming and costly [21]. Unfortunately, in this study, no data were available on patients' place of residence, which could be considered in future research.