CFM : comprehensive fuzzy logic-based modelling algorithm in breast cancer prognosis

Abstract

A novel supervised machine learning algorithm that improves the 1993 Sugeno-Yasukawa (SY) modelling algorithm to produce fuzzy models is presented in this thesis. There have been over 1500 citations of their study, however, there are only a handful of studies done on improving the performance of the SY technique. A modified version of SY modelling algorithm (CFM: Comprehensive Fuzzy Modelling) which produces statistically significant improvements of the original work of SY with regards to classification accuracy is introduced in this thesis. There are several reasons behind the shortcomings of the original work. Firstly, although the modelling technique was novel at the time it was presented, some of the component algorithms used in the technique were basic and now need improvements. Secondly, the algorithm developed for the fine tuning of the initially generated fuzzy models operates in a serial manner. In other words, it fine tunes different parameters of the model one at a time. This is not an effective method compared to parallel fine tuning performed by other algorithms such as a genetic algorithm. Lastly, the SY technique creates a group of intermediate models before creating the final model. In the SY technique, unlike the presented technique, there is no fine tuning of the intermediate models that are the building blocks of the final fuzzy model of the system. By overcoming these problems, the CFM technique achieves better performance. Simulations were performed for benchmark machine learning datasets to evaluate the modelling algorithm. The classification accuracy of the CFM algorithm for these datasets were among the highest compared to the existing studies and improved the SY algorithm results.

As an application area, we have focused on patients' survival classification diagnosed to be affected by breast cancer. For women, breast cancer is the most common cancer and has the highest mortality rate based on a recent survey performed by the Australian Institute of Health and Welfare. Therefore the analysis of prognosis factors and survival classification of breast cancer are of great importance. Two case studies on a breast cancer treatment dataset provided by the Canberra Hospital were performed by employing the CFM algorithm developed in this thesis. This contains over 10 years of follow-up data for 814 breast cancer patients. For these two case studies, the presented fuzzy logic based technique and supervised machine learning techniques were employed for classifying patients into survival categories. In both case studies, the CFM algorithm achieved the highest classification accuracy for survival classification compared to the other machine learning algorithm. Show more