Development of a Bayesian Network for Diagnosis of Breast Cancer

This paper describes the design, implementation, and preliminary evaluation of a Bayesian network that models the presence or absence of breast cancer based on demographic risk factors, mammographic signs, and physical findings. The network is implemented using the Bayesian Network Generation system (BNG) [11], which generates Bayesian networks from knowledge bases consisting of first-order probability logic sentences. Given a knowledge base of rules, a set of evidence E, and a query Q, the BNG system generates the structurally minimal network to compute P(Q | E). (A network is structurally minimal if it contains no barren nodes and no nodes d-separated from the query.) The generated network is passed to the IDEAL system [22] to perform the probabilistic computation. 2 Problem Overview Mammography is an important tool in early detection of breast cancer. Indeterminate mammographic findings often challenge the physician to distinguish between a malignant or benign condition. Successful diagnosis depends on the ability of a physician to recognize and evaluate a mammographic abnormality, as well as integrate the information from multiple clinical aspects (e.g., risk factors, physical findings) to determine the likelihood of breast cancer [9]. 2 2.1 Risk Factors and Physical Symptoms Recent demographic investigations as described by Gail et al. [10] and Colditz [4] have reported several risk factors considered to increase a woman's chance of developing breast cancer. Breast cancer incidence rates increase with age. Breast cancer is uncommon for a woman less than age 40, though this group shows the fastest rate of increase of the disease. Positive diagnoses occur with greater frequency in post-menopausal women. An early menarche, with its associated early onset of regular ..