Abstract
The Biofield Diagnostic System (BDS) uses a score formed with measured skin surface electropotentials and a prior Level Of Suspicion (LOS) value (predicted by the physician based on the patient's ultrasound or mammography results) to calculate a revised Post-BDS LOS to indicate the presence of breast cancer. The demographic details, BDS test results, and the recorded electropotential values form a potentially useful dataset, which can be further explored with data mining tools to extract important information that can be used to improve the current predictive accuracy of the device. According to the proposed data mining framework, the BDS dataset with 291 cases was first pre-processed to remove outliers and then used to select relevant and informative features for classifier development and finally to evaluate the capability of the built classifiers in detecting the presence of the disease. Two popular feature selection techniques, namely, the filter and wrapper methods, were used in parallel for feature selection. A few statistical inference based classifiers and neural networks were used for classification. The proposed technique significantly improved the BDS prediction accuracy. Also, the use of prior LOS and, hence, the Post-BDS LOS, associates a mild subjective interpretation to the current prediction methodology used by BDS. However, the feature subset selected in our analysis that gave the best accuracy did not use either of these features. This result indicates the possibility of using BDS as a better objective assessment tool for breast cancer detection.