A Method For Predicting The Characteristics Of Cerebral Cortex Thickness Based On Early Prediction Of Alzheimer's Disease

Alzheimer’s disease is a neurodegenerative disease of the brain, It is called senile dementia in our daily life. Mild cognitive impairment is generally referred to as the precursor symptoms of Alzheimer’s disease. Related research found that approximately 10 percent to 15 percent of MCI patients are transformed into AD patients. It is only 2 percent for the normal controls. Undoubtedly, MCI patients are at higher risks. According to the literature, AD is a kind of hidden and imperceptible neurological disease, it is difficult to be discover on its early stage. Thus, for the sake of early diagnosis of AD progression, it is extremely urgent to delay and prevent such conversion, explore the efficient means of predicting the progressive tendency. However, so far there are no well-recognized methodologies of detecting the progressive process and accurately forecasting such trend. In the present study, we presented a novel predictive method based on distinguishing stable MCI(sMCI, who remain stable in a period of time) from AD, receiving relatively better performance than commonly used and providing supportive and new thinking of early diagnosis and intervention of AD.Based on the brain morphology analysis, more and more scholars suggested that biomarkers were extracted from cerebral cortex could get excellent performance in classification tasks. However, how to choose the optimal set of features is still hard to determine. Since feature selection methods are varied, and a satisfied method might get much better result than others. In addition, the performance based on cortical thickness still has space for improvement. In this article, cortical thickness data was extracted from magnetic resonance imaging(MRI) as the input feature to realize AD early diagnosis. The main contributions of this article are as follows:(1) The Rank combination method combines the mutual information and Pearson correlation coefficient theory, not only considers the 78 cerebral cortex thickness in correlation with class label, also considers the selected features redundancy between each other in the process of feature selection. Avoide only considering relevance and ignore features redundancy eventually lead to larger redundancy between features, or only consider the redundancy between features and ignore each specificity of class label result in uncertainty, which leads to the final classification result is not ideal.(2) Compared with mRMR,which just combines two ranking systems, and which one is better was not considered. In order to solve the above problem, two kinds of fusion system with the corresponding weight coefficient in this paper. By experiments,we can find the best feature subset. Compared with other methods, we achieved a higher accuracy of classification.(3) An experiment combined the proposed feature selection method and support vector machine(SVM), when the result achieved the highest accuracy,we selected some areas in this paper. By contrast with brain images, the characteristics of brain area mainly concentrated in the parahippocampal gyrus and cingulate, which is similar to previous studies. At the same time, we also found different brain regions with previous research(the forehead back and olfactory cortex), which provides important reference for the further research.