An optimum ANN-based breast cancer diagnosis: Bridging gaps between ANN learning and decision-making goals

Abstract

It is difficult to overestimate the importance of appropriate breast cancer diagnosis, as the disease ranks second among all cancers that lead to death in women. Many efforts propose data analytic tools that succeed in predicting breast cancer with high accuracy; the literature is abundant with studies that report close-to-perfect prediction rates. This paper shifts the focus of improvement from higher accuracy towards better decision-making. Quantitatively, we have shown more accuracy does not always lead to better decisions, and the process of Artificial Neural Networks (ANN) learning can benefit from the inculcation of decision-making goals. We have proposed a decision-oriented ANN classification method called Life-Sensitive Self-Organizing Error-Driven (LS-SOED), which enhances ANN’s performance in decision-making. LS-SOED combines the supervised and unsupervised learning power of ANN to handle the inconclusive nature of hidden patterns in the data in such way that the best possible decisions are made, i.e. the least misclassification cost (the minimum possible loosing of life) is achieved. The learning power of SOED matches, if not excels, the best performances reported in the literature when the objective is to achieve the highest accuracy. When the objective is to minimize misclassification costs, we have shown, on average, in one dataset more than 30 years of life for a group of 283 people, and in another more than 8 years of life for a group of 57 people can be saved collectively.