Classification Model For Super Early Stage Of Alzheimer's Disease Based On Support Vector Machine And Multi-modal Information Fusion

AD(Alzheimer's disease,AD),mainly featuring cognitive impairment clinically,is a neurodegenerative disease with slow onset When obvious symptoms appear,patients are often in the middle and late stages and cannot be effectively treated.Therefore,early diagnosis and intervention treatment are important for effective control of the AD.The patient's brain cell function remains well in the SCD stage and this stage is within the optimal time window for AD treatment.Therefore,the effective and accurate classification aimed at the pre-AD stage(ie MCI(Mild Cognitive Impairment,MCI)and SCD(Subjective Cognitive Decline,SCD)stage)has received more and more attention in recent years.In addition,most of the current classification studies for MCI,SCD,and AD only consider a single biomarker,and lack the fusion of different biomarkers to provide sufficient information for accurate classification and diagnosis.Therefore,this thesis proposes a preclinical classification algorithm for AD based on multi-modal SVM(Support Vector Machine,SVM).This algorithm uses multi-modal medical image information to explore the brain mechanism of AD occurrence from multiple angles,and effectively improve the accuracy of the classification model by fusing gray matter volume information extracted from sMRI(Structural Magnetic Resonance Imaging sMRI)and extracted ALFF(Amplitude of Low Frequency Fluctuation,ALFF)from rsfMRI and structural connection network.extracted from DTI(Diffusion Tensor Imaging,DTI).In view of the problem that the brain structural and functional biomarkers of patients in SCD stage are not significantly changed,the brain area with significant differences between the AD group and the normal control group is used as a template.The template is used to exclude irrelevant features,which can effectively avoid the slightly structural and functional changes of the SCD stage being missed.Comparing the advantages and disadvantages of feature selection algorithms such as f_classif,ReliefF,and recursive feature elimination algorithm,use the obtained optimal feature subset to train the classification model.Then using the comple-mentarity between different modal data to design a classification model based on multi-modal S VM,and comparing the multi-modal S VM with single-modal S VM and multiple kernel SVM based on weighted summation kernel.Experiments show that the proposed method has an accuracy rate of 86.67%when classifying between SCD and normal controls,which is 13.34%higher than that of sMRI-based single-mode SVM classification,and 20%higher than that of rs-fMRI-based single-mode SVM classification,13.34%higher than DTI-based single-mode SVM classification accuracy,and 6.67%higher than multiple kernel SVM based on weighted summing kernels;at the same time,for aMCI,AD clas-sification,compared with single-modal SVM classification,the classification accuracy of multi-mode SVM has also been significantly improved.In addition,the method proposed in this thesis has a short running time,is easy to embed in the traditional SVM classifier,and has great application value.Through the diagnosis of SCD,intervention and treatment can be carried out in the early stage of this disease,which has great clinical significance.