سفارش ویژگی های تکاملی در شبکه های بیزی برای پیش بینی سندرم متابولیک

The metabolic syndrome is a set of risk factors that include abdominal obesity, insulin resistance, dyslipidemia and hypertension. It has affected around 25% of adults in the US and become a serious problem in Asian countries recently due to the change in dietary habit and life style. On the other hand, Bayesian networks that are the models to solve the problems of uncertainty provide a robust and transparent formalism for probabilistic modeling, so they have been used as a method for diagnostic or prognostic model in medical domain. Since the K2 algorithm, a well-known algorithm for Bayesian networks structure learning, is influenced by an input order of the attributes, an optimization of BN attribute ordering has been studied as a research issue. This paper proposes a novel ordering optimization method using a genetic algorithm based on medical expert knowledge in order to solve this problem. For experiments, we use the dataset examined twice in 1993 and 1995 in Yonchon County of Korea. It has 18 attributes of 1193 subjects participated in both surveys. Using this dataset, we make the prognostic model of the metabolic syndrome using Bayesian networks with an optimized ordering by evolutionary approach. Through an ordering optimization, the prognostic model of higher performance is constructed, and the optimized Bayesian network model by the proposed method outperforms the conventional BN model as well as neural networks and k-nearest neighbors. Finally, we present the application program using the prognostic model of the metabolic syndrome in order to show the usefulness of the proposed method.

The metabolic syndrome is composed of a cluster of metabolic disorders including abdominal obesity, insulin resistance, dyslipedemia and hypertension, and the correlation between metabolic syndrome and other diseases such as diabetes and coronary heart disease is reported in the literature (Cabre et al., 2008 and Lee et al., 2009). It affects around 25% of adults over the age of 20 and up to 45% over age 50 in the United States (Mehta & Reilly, 2004). These days, it is found even in children and adolesecents as a ratio of approximately 4%–7% (Pan & Pratt, 2008). In Asian countries, it has become a significant problem lately due to the change in dietary habit and life style (Moon et al., 2003 and Son et al., 2005). In situations like this, many groups have been studying the metabolic syndrome from all over the world (Cabre et al., 2008, Lee et al., 2009, Mehta and Reilly, 2004, Moon et al., 2003, Pan and Pratt, 2008 and Son et al., 2005). Recently, computer based health-care systems have been studied and developed a lot according to an advancement of artificial intelligence techniques such as image processing and expert systems and increase of people’s concern for their health (Herzbert et al., 2009). The systems for computer-based diagnosis or pervasive health-care are the representative examples (Innocent and John, 2004 and Osmani et al., 2008). The Bayesian network, one of the AI techniques, has emerged in recent years as a powerful technique for handling uncertainty in complex domains (Larranaga, Poza, Yurranmendi, Murga, & Kuijpers, 1996). It is a model of a joint probability distribution over a set of random variables. The Bayesian network is represented as a directed acyclic graph where nodes correspond to variables and arcs correspond to probabilistic dependencies between connected nodes (Chen & Blanchette, 2007). Bayesian networks have been used for prediction or classification problem in the medical domain and shown high performance. In particular, they have been applied successfully to the modeling of diagnosis and prognosis for diverse diseases (Antal et al., 2004, Antal et al., 2003, Aronsky and Haug, 2000, Charitos et al., 2009, Gerven et al., 2007, Getoor et al., 2004, Maskery et al., 2008, Sierra et al., 2001, Tucker et al., 2005 and Wang et al., 1999). There have been many black box tools that classify or predict several diseases, and neural networks are the representative example. Bayesian networks have strengths that they can use the domain knowledge easily and analyze the results compared to them (Wang et al., 1999). Even though they are sometimes not better than neural networks in terms of accuracies, Bayesian networks are appropriate methods in the medical domain that requires to analyze the results with medical knowledge. This paper used a genetic algorithm to solve an optimization problem in medical prognostic modeling. Specifically, this paper deals with a problem that predicts the metabolic syndrome with the dataset obtained in Yonchon County of Korea. This paper constructs a prognostic model using Bayesian network, and has used the K2 algorithm by Cooper and Herskovits in order to learn its structure (Larranaga, Poza, et al., 1996). Since the result of the K2 algorithm is influenced by an input ordering of the attributes, an optimization of this ordering has been also studied (Hruschka and Ebecken, 2007, Hruschka et al., 2007, Hsu, 2004, Larranaga et al., 1996 and Song et al., 2007). This paper proposes an efficient optimization method using medical domain knowledge and a genetic algorithm in order to solve this problem. Different from the conventional methods, after clustering similar attributes into each group, an ordering of the groups and an ordering of the attributes in each group have been performed in turns. As applying the medical domain knowledge, an efficient and reliable modeling has been conducted. Subsequently, the experiments using the proposed prognostic model have been conducted after the structure and parameter learning processes, and an application program using this model has been presented in order to show its usefulness. This paper is organized as follows. Section 2 presents the definition of the metabolic syndrome, the advantages of BN application to medical domains and related works on Bayesian network attribute ordering. In Section 3, the proposed method using medical domain knowledge and the genetic algorithm is described. In Section 4, various experimental results in order to show the usefulness of the proposed method are provided, and Section 5 concludes the paper with summary.

This paper proposed the Bayesian network model in order to predict the metabolic syndrome. In processes of building the prognostic model, we applied the medical domain knowledge to make the model more reliable. We also applied the genetic algorithm to optimize attribute ordering, and we completed the model more efficiently using the medical domain knowledge in this optimization process. We confirmed that the proposed method provided better performance comparing with the model before ordering optimization and other models such as neural network and k-nearest neighbor. Finally, we presented a useful application program using the prognostic model built by the proposed method. For future work, we will apply the proposed method to predict other diseases. Also, the application program can be used in mobile devices like a smart phone. With the use of ubiquitous sensors and automatical collection of user information, it can contribute to implement pervasive health-care systems.