Mutual information-based Fisher discriminant analysis for feature extraction and recognition with applications to medical diagnosis

This paper presents a novel discriminant analysis (DA) for feature extraction using mutual information (MI) and Fisher discriminant analysis (MI-FDA). Most DA algorithms for feature extraction are based on a transformation which maximizes the between-class scatter and minimizes the within-class scatter. In contrast, the proposed method uses the Fisher's criterion to find a transformation that maximizes the MI between the transferred features and the target classes and minimizes the redundancy. The performance of the proposed method is evaluated using UCI databases and compared with the performance of some DA-based algorithms. The results indicate that MI-FDA provides a robust performance over different data sets with different characteristics. On average, an accuracy rate of 81.3% was achieved using MI-FDA.