A Hybrid Bayesian Network Model for Predicting Breast Cancer Prognosis.
10.4258/jksmi.2009.15.1.49
- Author:
Jong Pill CHOI
1
;
Tae Hwa HAN
;
Rae Woong PARK
Author Information
1. Department of Medical Informatics, Ajou University School of Medicine, Korea. veritas@ajou.ac.kr
- Publication Type:Original Article
- Keywords:
Bayesian Network;
Neural Network;
Prognosis;
Breast Cancer
- MeSH:
Breast Neoplasms*;
Classification;
Decision Trees;
Female;
Humans;
Logistic Models;
Neural Networks (Computer);
Prognosis*;
SEER Program;
Sensitivity and Specificity
- From:Journal of Korean Society of Medical Informatics
2009;15(1):49-57
- CountryRepublic of Korea
- Language:English
-
Abstract:
OBJECTIVE: Breast cancer is one of the most common cancers affecting women. Both physicians and patients have concerned about breast cancer survivability. Many researchers have studied the breast cancer survivability applying artificial nerural network model (ANN). Usually ANN model outperformed in classification of breast cancer survivability than other models such as logistic regression, Bayesian network (BN), or decision tree models. However, physicians in the fields hesitate to use ANN model, because ANN is a black-box model, and hard to explain the classification result to patients. In this study, we proposed a hybrid model with a degree of the accuracy and interpretation by combining the ANN for accuracy and BN for interpretation. METHODS: We developed an artificial neural network, a Bayesian network, and a hybrid Bayesian network model to predict breast cancer prognosis. The hybrid model combined the artificial neural network and the Bayesian network to obtain a good estimation of prognosis as well as a good explanation of the results. The National Cancer Institute's SEER program public-use data (1973-2003) were used to construct and evaluate the proposed models. Nine variables, which are clinically acceptable, were selected for input to the proposed models' nodes. A confidence value of the neural network served as an additional input node to the hybrid Bayesian network model. Ten iterations of random subsampling were performed to evaluate performance of the models. RESULTS: The hybrid BN model achieved the highest area under the curve value of 0.935, whereas the corresponding values of the neural network and Bayesian network were 0.930 and 0.813, respectively. The neural network model achieved the highest prediction accuracy of 88.8% with a sensitivity of 93.7% and a specificity of 85.4%. The hybrid Bayesian network model achieved a prediction accuracy of 87.2% with a sensitivity of 93.3% and a specificity of 83.1%. The results of the hybrid Bayesian network model were very similar to the neural network model. CONCLUSION: In the experiments, the hybrid model and the ANN model outperformed the Bayesian network model. The proposed hybrid BN model for breast cancer prognosis predictin may be useful for clinicians in the medical fields, as the model provides both high degree of performance inherited from ANN and good explanation power from BN.
