A coupled feature representation based MRI biomarker for Alzheimer's disease prediction

摘要

Mild Cognitive Impairment (MCI) is an inevitable stage before patient convert to Alzheimer's disease (AD). In MCI cases, only part will further progress to AD. This progress is irreversible and highly lethal. Based on that fact, precisely predict who will convert to AD could allow the MCI patients to receive timely treatment for delaying the onset of symptoms, and this also highly helpful to the clinical trials research. In this work, we proposed a novel biomarker based on Coupled Feature Representation of Magnetic Resonance (MR) images for the AD conversion prediction. First, we adopted a Taylorlike expansion based Coupled Feature Representation method to obtain the linear and nonlinear information between the regional features. Then a support vector machine (SVM) classifier was trained to evaluate decision values of each sample, which we call the coupling biomarker. Ultimately, the coupling biomarker is connected with age and cognitive measures(CM) to obtain the final prediction. In the experimental results, the proposed coupling biomarker achieved a 10-fold cross-validated (CV) Accuracy (ACC) of 75.4%, and area under the receiver operating characteristic curve (AUC) of 81.7% in prediction task for MCI subjects, and the combined biomarker further achieved a 10-fold CV ACC of 84.4% and AUC of 92.0%. In the last, the joint biomarker which merge the coupling and grading biomarkers achieved a 10-fold cross-validated ACC of 85.7% and AUC of 92.6%, and both SPE and SEN higher than 80.0%. The results presented in this study demonstrate a significant contribution in accurate prediction and also the potential and expandable of the framework for prediction of conversion from MCI to AD.