سیستم پشتیبانی تصمیم برای مدیریت تحت درمان با وارفارین با استفاده از شبکه های بیزی

Warfarin therapy is known as a complex process because of the variation in the patients' response. Failure to deal with such variation may lead to death as a result of thrombosis or bleeding. The possible sources of variation such as concomitant illnesses and drug interactions have to be investigated by the clinician in order to deal with the variation. This paper describes a decision support system (DSS) using Bayesian networks for assisting clinicians to make better decisions in Warfarin therapy management. The DSS is developed in collaboration with a Swedish hospital group that manages Warfarin therapy for more than 3000 patients. The proposed model can assist the clinician in making dose-adjustment and follow-up interval decisions, investigating variation causes, and evaluating bleeding and thrombosis risks related to therapy. The model is built upon previous findings from medical literature, the knowledge of domain experts, and large dataset of patients.

Warfarin is an oral anticoagulant that is mainly used for preventing thrombosis and embolism in several clinical disorders including atrial fibrillation and pulmonary embolism. The duration of Warfarin therapy is often between three months and a lifetime [12]. The effects of Warfarin are generally monitored by International Normalized Ratio (INR) which is the ratio of patient's blood coagulation time to a reference sample. Patients' INR values should be kept within a target therapeutic range since mortality risk will increase considerably if the INR value is outside this range [27]. According to a large multicenter randomized study by Poller et al. [28] the quality of the manual Warfarin management is still not adequate as patients can be kept within therapeutic range only 65% of the time. In Sweden, approximately 150,000 patients are treated with Warfarin and this number is increasing steadily [35]. According to Swedish statistics, around 12 patients die each year as a result of bleedings caused by Warfarin therapy and similar anticoagulant therapies [34]. There appears to be a significant room for improvement to keep the patients' INR values within therapeutic range and to minimize Warfarin related risks. There are a number of ‘variation factors’ such as drug interactions and concomitant diseases which can increase Warfarin therapy risks by causing unexpected increases or decreases in the INR value [12]. It is difficult to know the presence of these factors in advance since some of the variation factors are commonly consumed products such as leafy vegetables, and their consumption usually varies in time. Clinicians have to investigate the presence of these factors in order to lower the risks of Warfarin therapy. Decision support systems (DSS) have been used for assisting the decision making in Warfarin therapy since 1976 [40]. The main outputs of the Warfarin DSS are dose adjustments (how much the dose should be adjusted?) and follow-up intervals (when should the patient take the next INR test?) [8]. Various studies have shown that DSS are capable of increasing the quality of Warfarin therapy [10] and [28]. None of the reviewed DSS assists the investigation of variation factors during Warfarin therapy. Moreover, many existing DSSs such as regression models do not deal with the dynamic nature of the therapy (See Section 2). This paper proposes a DSS using Bayesian networks (BN) for assisting the management of Warfarin therapy. The objective of this DSS is to support dose and follow-up interval decisions while predicting cerebral bleeding and stroke risks, and to assist the investigation of variation factors. Other advantages of the proposed DSS include its flexibility in terms of inputs and outputs, and training support for clinicians. The BN model has been built in collaboration with the Skaraborg Hospital Group, SkaS, Sweden. The structure of the model is based on relevant medical findings published in reputable international journals and the knowledge of the physicians and nurses who are actively working on Warfarin therapy. The parameters of the model are identified from a large dataset of patients, elicitations with the domain experts and published statistics in medical literature. Genie-SMILE software [11] was used for building and calculating the BN. The rest of this paper is organized as follows. Previous DSS for Warfarin therapy are reviewed in Section 2. The BN model for Warfarin therapy management is described in Section 3. Validation of the model is presented in Section 4. An example of using the model is provided in Section 5. Discussions and conclusions are presented in 6 and 7 respectively.

The aim of the work described here is to propose a DSS using BNs for assisting clinicians in managing Warfarin therapy. The BN model is built upon the medical literature, the data and the knowledge of domain experts from a Swedish hospital. The model is capable of recommending dose adjustment and follow-up interval decisions. Unlike most of other DSS utilized in the Warfarin therapy, the BN model also assists investigation of the factors that may cause variation in INR and evaluates the cerebral bleeding and stroke risks related to the therapy. In order to validate the dose adjustment and follow-up interval decisions of the model, 10-fold cross validation method was used. It was found that the actual dose adjustment and follow-up interval decision was one of the two most probable outputs of the model in 84% and 79% of the cases respectively. The BN model's performance in investigating possible variation factors was tested with the experts, as there was no data available for this validation. Out of the 30 cases generated for this test, the experts agreed with the model in 83% of the cases. The part of the model used to estimate the cerebral bleeding and stroke risks is built upon the validated statistics in the medical literature; however, since no data about bleeding and stroke incidences were available, no additional validation was possible for this part of the model. For further research, this model could be expanded to manage all of the Warfarin therapy including both the initiation phase and the maintenance phase. Moreover, it could be adapted to manage other target therapeutic ranges such as 2.5 to 3.5. More conclusive statements about the model's reliability could be made if the model were validated prospectively. Finally, the model's user interface can be simplified to make it easier to use and update by the clinicians.