Noise and Outlier Filtering in Heterogeneous Medical Data Sources

Abstract

There is a growing interest in studying the common features from multiple data sources. Fusing information from multiple heterogenous data sources promises to identify complex multivariate relationships among the heterogeneous sources. Such relationships can provide additional connectivity across the sources. A common way to analyze the relationships between a pair of data sources based on their correlation is canonical correlation analysis (CCA). CCA seeks for linear combinations of all variables from each dataset with maximal correlation between the two linear combinations. However, the existence of non-informative data points and features makes it challenging for CCA to identify significant relationships among the examined datasets. In this paper, we propose a novel method, NORA, Noise-Outliers Removal Algorithm, that can be used to filter out the non-informative data points and features before applying the CCA. NORA was applied to preprocess two epilepsy modalities, the MRI and neuropsychology, prior to applying CCA to find the association between them. The results show that the proposed method leads to interpretable results when noise plays a significant role in the acquisition of the data.