Research On Early Diagnosis Algorithm Of Alzheimer's Disease Based On Nuclear Machine Learning

Previous studies have shown that in computer-aided diagnosis of AD,the available sample size is usually small,but the feature dimension is higher,and the relationship between intra-group features and inter-group features is complex.However,the traditional machine learning methods do not solve these problems,which leads to misjudgment and low accuracy.In response to these problems,this thesis focuses on how to combine kernel learning with traditional machine learning methods to achieve a better feature selection effect and higher classification accuracy.Firstly,a AD diagnostic method based on KPCA-FDA and KSVM is proposed.The kernel method is combined with PCA,FDA,and SVM to improve the classification performance of AD.Specifically,first,the original feature subset is placed into a KPCA module,and then the original data set is projected onto a higher dimensional kernel space,the principal component coefficients are reduced to increase linear separability.Then,project the KPCA coefficients into a more efficient FDA to select the optimal feature subset.Finally,the KSVM combined with the new feature subset to classify the data of AD,MCI and NC,so as to achieve a better classification effect.The sMRI data of the subjects are used for the experiment.The results showed that the method have a good classification performance in distinguishing AD and NC.The classification accuracy is 92.34%,the sensitivity is 91.71%,the specificity is 90.04%,and the AUC is 0.9143.When distinguishing between MCI and NC the classification accuracy is 75.49%,the sensitivity is 80.45%,the specificity is 70.23%,and the AUC is 0.8036.Secondly,a multi-modal AD early diagnosis algorithm based on MK-SVM was proposed.This method combines sMRI and PET data to obtain more auxiliary information to improve the classification performance of AD.First,the FDA,Kernel method and LPP algorithm are combined to generate a KLFDA algorithm to reduce the dimension of the extracted sMRI and PET data features.Then,two MKL based multivariate classification methods,WTA-KSVM and MWV-KSVM,are proposed to classify the data of AD,MCI and NC.Finally,the data of sMRI and PET are used to verify the performance of the classification algorithm.The results show that the classification results obtained by the two multivariate classification methods on multimodal data(sMRI+PET)are better than those obtained on single modal data(sMRI or PET).The WTA-KSVM method achieved the best classification performance,the classification accuracy is80.41%,the sensitivity is 81.56%,the specificity is 80.58%,the AUC is 0.8175.The classification accuracy of MWV-KSVM is 80.07%,the sensitivity is 81.23%,specificity is80.16%,and AUC is 0.8114.Compared with other methods,the two multivariate classification methods proposed in this paper have achieved better classification performance.Based on the kernel method and machine learning,this thesis makes a preliminary exploration and research on the early diagnosis of AD.Two early diagnosis algorithms for AD based on kernel machine learning are proposed,which provide some ideas and directions for future generations.