A reliable neural network-based decision support system for breast cancer prediction

Abstract

Axillary lymph node (ALN) metastasis status is an important prognostic marker in breast cancer and is widely employed for tumour staging and defining an adjuvant therapy. In an attempt to avoid invasive procedures which are currently employed for the diagnosis of nodal metastasis, several markers have been identified and tested for the prediction of ALN metastasis status in recent years. However, the nonlinear and complex relationship between these markers and nodal status has inhibited the effectiveness of conventional statistical methods as classification tools for diagnosing metastasis to ALNs. The aim of this study is to propose a reliable artificial neural network (ANN) based decision support system for ALN metastasis status prediction. ANNs have been chosen in this study for their special characteristics including nonlinear modelling, robustness to inter-class variability and having adaptable weights which makes them suitable for data driven analysis without making any prior assumptions about the underlying data distributions. To achieve this aim, the probabilistic neural network (PNN) evaluated with the .632 bootstrap is investigated and proposed as an effective and reliable tool for prediction of ALN metastasis. For this purpose, results are compared with the multilayer perceptron (MLP) neural network and two network evaluation methods: holdout and cross validation (CV). A set of six markers have been identified and analysed in detail for this purpose. These markers include tumour size, oestrogen receptor (ER), progesterone receptor (PR), p53, Ki-67 and age. The outcome of each patient is defined as metastasis or non-metastasis, diagnosed by surgery. This study makes three contributions: firstly it suggests the application of the PNN as a classifier for predicting the ALN metastasis, secondly it proposes a the .632 bootstrap evaluation of the ANN outcome, as a reliable tool for the purpose of ALN status prediction, and thirdly it proposes a novel set of markers for accurately predicting the state of nodal metastasis in breast cancer. Results reveal that PNN provides better sensitivity, specificity and accuracy in most marker combinations compared to MLP. The results of evaluation methods’ comparison demonstrate the high variability and the existence of outliers when using the holdout and 5-fold CV methods. This variability is reduced when using the .632 bootstrap. The best prediction accuracy, obtained by combining ER, p53, Ki-67 and age was 69% while tumour size and p53 were the most significant individual markers. The classification accuracy of this panel of markers emphasises their potential for predicting nodal spread in individual patients. This approach could significantly reduce the need for invasive procedures, and reduce post-operative stress and morbidity. Moreover, it can reduce the time lag between investigation and decision making in patient management.