An artificial neural network based feature evaluation index for the assessment of clinical factors in breast cancer survival analysis

This study aims to identify the most and least significant prognostic factors for breast cancer survival analysis by means of feature evaluation indices derived from multilayer feedforward backpropagation neural networks (MLFFBPNN), fuzzy k-nearest neighbour classifier (FK-NN) and a logistic regression-based backward stepwise method (ER). The data used for the survival analysis were collected from 100 women who had been clinically diagnosed with breast disease in the form of carcinoma or benign conditions. The data set consists of seven different histological and cytological prognostic factors and two corresponding outputs to be predicted (whether the patient is alive or dead within 5 years of diagnosis). The MLFFBPNN, FK-NN and LR based indices identified different subsets of the factors as the most significant sets. We therefore suggest that it could be dangerous to rely on one method's outcome for assessment of such factors. It should also be noted that "S-phase fraction" (SPF) is the common cytological factor identified by all three methods while none of the three methods identified another cytological factor, namely "minimum (start) nuclear pleomorphism index" (NPImin). We, therefore, conclude that "S-phase fraction" and "minimum (start) nuclear pleomorphism index" appear to be the most and least important prognostic factors, respectively, for survival analysis in breast cancer patients, and should be investigated thoroughly in future clinical studies in oncology.