Classification of Alzheimer’s disease and Parkinson’s disease using a support vector machine and probabilistic outputs

Abstract

Alzheimer's disease (AD) and Parkinson's disease (PD) are both prominent central nervous system diseases that are frequently diagnosed and studied using brain single-photon emission computed tomography (SPECT). Owing to divergent clinical features, AD and PD are often considered distinct diseases; however, it is difficult to distinguish AD from PD on SPECT. Tools for objectively analyzing differences between AD and PD on SPECT images are not currently available. To construct a model for discriminating AD from PD in Japanese patients, we used a support vector machine (SVM) and SPECT images acquired at two different time points after radiotracer injection to extract the determinant regions for classification. We assessed SPECT images from 68 Japanese patients with AD or PD. After pre-processing noise voxels, a non-linear SVM classification with Gaussian kernels was adopted to construct the predictive model. The best SVM model was highly accurate for distinguishing AD from PD. The accuracy of this model was 98.1% for leave-one-out cross-validation and 78.6% for the test set. Our data showed that the temporal, sub-lobar, parietal, limbic, and frontal areas exhibited decreased regional cerebral blood flow in AD; whereas the frontal, anterior, parietal, and occipital areas exhibited decreased regional cerebral blood flow in PD. Here, we present a useful SVM model for classifying AD versus PD using SPECT images and show the utility of two-time-point SPECT imaging for AD/PD discrimination.