A Novel Hybrid Classifiers based Model for mining in Neuro-imaging

Resting state functional magnetic resonance imaging (fMRI) has great clinical significance in detection and diagnosis of epilepsy. Various neuro-imaging markers have been extracted from time series data using statistical and connectivity measures. Powerful data mining rules, association based techniques and classifiers are used to extract information from big datasets. The application of data mining in neuro-imaging has been explored rarely. This paper proposes a hybrid classification based ensemble learning method for mining in neuro-imaging to detect epilepsy. The paper combines various feature extraction and feature selection methods to extract discriminative statistical, evolutionary and functional connectivity features at multiple levels. The features are then presented to a hybrid classification system that utilizes ensemble learning methods to train the classifiers and the best classification result is selected based on majority voting. The extraction of bio-markers from the neuro-imaging system has been explained and the combination of feature vectors and their impact on classification accuracy has been explained. The classification accuracy results using some of the biomarkers have been very encouraging. The proposed methodology combines the linear, non-linear and probabilistic aspects of the dataset and can be extended to any neuro-imaging system for clinical diagnosis and prognosis.