Hide/Show Apps

UTADIS based multi-objective evolutionary algorithms for medical diagnosis problems

Download
2019
Mahmutoğulları, Halenur Şahin
We develop hybrid methods that integrate multi-criteria decision making, evolutionary algorithms and machine learning to be used in medical diagnosis problems. The proposed models classify patients into two categories according to their disease status with the aim of obtaining high classification performances both classes under consideration. First, we develop a Mixed-Integer Linear Programming approach, Parametrized Classification Model (PCM), which is based on UTADIS. By solving PCM multiple times with various values of a specific parameter, we obtain a set of solutions spread over the Pareto-optimal front in the space of true positive and true negative responses. Then, to combine strong aspects of these solutions, we integrate PCM with evolutionary algorithms, NSGA-II and RECGA, to tune the classification parameters acquired by PCM. NSGA-II favors non-dominated solutions in terms of sensitivity and specificity and RECGA aims to perform well particularly in situations where the incidence of the disease may be relatively low, such as general screening. We call the developed integrated models as PCM+NSGA-II and PCM+RECGA, respectively. In order to observe the model performances, we try them with three different datasets which are about coronary stent patients and breast cancer. Furthermore, we apply several well-known machine learning algorithms to these datasets and compare the results with the results of PCM+NSGA-II and PCM+RECGA. Additionally, for the coronary stent dataset, the model performances are compared with those of cardiologists. The results indicate that PCM+NSGA-II and PCM+RECGA are promising classification algorithms that can be used in medical decision support tools by medical experts.