Feature reduction in heterogeneous data sets via sequential search techniques

In the clinical setting, pattern recognition techniques promise significant improvement for disease state detection. However, many medical conditions are not reliably described by a single defining feature. In such heterogeneous groups, capturing synergistic effects among features becomes vital to classification success. Additionally, many commonly employed feature selection techniques, such as statistical tests, may not work well when applied to problems involving heterogeneous groups. This treatise explores the application of Sequential Forward Floating Search (SFFS) techniques in the high-dimensional data sets common to the neuroimaging field. Using both multimodal neuroimaging data and calibrated synthetic data sets, the project seeks to characterize SFFS’s effectiveness versus established statistical approaches common in the literature.