Data mining technique for identification of diagnostic biomarker to predict Schizophrenia disorder

• Published in: 2014 IEEE International Conference on Computational Intelligence and Computing Research pp. 1 - 8
• Main Authors: GeethaRamani, R., Sivaselvi, K.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2014
• Subjects: Accuracy; Classification; Classification algorithms; Decision trees; Feature selection; Filtering; Filtering algorithms; Functional connectivity; Machine learning algorithms; Resting state fMRI; Schizophrenia; Vegetation
• ISBN: 1479939749; 9781479939749
• DOI: 10.1109/ICCIC.2014.7238525

In recent days, researchers are actively analysing the human brain to understand the underlying mechanism of heterogeneous psychiatric conditions. Schizophrenia is a severe neurological disorder which has been characterized by varying symptoms namely hallucinations, delusions and cognitive problems. In this paper, we have investigated the resting state fMRI images of 15 normal controls and 12 Schizophrenia patients by constructing functional connectome through image preprocessing techniques namely Realignment, temporal correction, filtering, etc., The parcellation of neuroimage is performed based on Automated Anatomical Labelling (AAL) atlas and 74 regions of interest (ROI) are identified. Functional connectome of each subject includes Pearson correlation values of mean time courses obtained between the regions. These region to region functional connectivity is considered as features and the feature selection technique namely Fisher filtering, ReliefF filtering and Runs filtering are applied. Then the features which are found by different filtering techniques are fed as input to the supervised non linear classifiers namely Random forest, C4.5, Cost sensitive classification and regression tree and K-Nearest Neighbour classification algorithm. These algorithms have produced classification rules which are used in the prediction of Schizophrenia disorder. C4.5 has achieved the higher predictive accuracy of 93% with leave-one out cross-validation and the predominant feature or diagnostic biomarker is obtained from the rule. This feature is one among the commonly identified feature of different feature selection techniques. The work has shown that the biomarker corresponds to the alterations in the functional connectivity between the brain regions namely Rolandic operculum and Postcentral gyrus of brain's left hemisphere which is involved in sensorimotor function of human.