Classifying Alzheimer’s Disease with Machine Learning via Wavelet Transform Subband Combinations

Abstract

The 3-D wavelet transform has been used with machine learning techniques to help identify Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI) through magnetic resonance imaging (MRI). Although this approach resulted in high accuracy, a large number of subband permutations were obtained due to the usage of the 3-D wavelet transform, which causes redundancy and computational inefficiency during classification. To address this issue, this study discovers the combination with the minimum subbands giving a comparable accuracy with the original approach that used all subbands. The Alzheimer’s Disease Neuroimaging Initiative (ADNI) was used to gain de-identified MRI scans for classification with 75% training and 25% test scans. The scans were standardized by pre-processing through skull stripping, segmentation, and smoothing using the MATLAB toolboxes Statistical Parametric Mapping 12 (SPM 12) and Computation Anatomy Toolbox 12 (CAT 12). Feature extraction was then completed through different 3-D wavelet transform subband combinations. Support vector machine (SVM) classification was used with radial basis function kernels (RBF) to screen patients for AD, MCI, or cognitively normal (CN) status 3 times per subband combination. The mean and standard deviation of the test accuracy for each combination were recorded. Among the tested combinations, the maximum mean test accuracy was 95.3%. To prevent overfitting, five-fold cross-validation was performed on the top 40% subband combinations based on the mean test accuracy. The maximum cross-validation accuracy is 97.3%. Therefore, this study shows the potential of using less 3-D wavelet transform subbands to help screen patients for AD or MCI in the future.