Domain transfer learning for MCI conversion prediction

Abstract

In recent studies of Alzheimer's disease (AD), it has increasing attentions in identifying mild cognitive impairment (MCI) converters (MCI-C) from MCI non-converters (MCI-NC). Note that MCI is a prodromal stage of AD, with possibility to convert to AD. Most traditional methods for MCI conversion prediction learn information only from MCI subjects (including MCI-C and MCI-NC), not from other related subjects, e.g., AD and normal controls (NC), which can actually aid the classification between MCI-C and MCI-NC. In this paper, we propose a novel domain-transfer learning method for MCI conversion prediction. Different from most existing methods, we classify MCI-C and MCI-NC with aid from the domain knowledge learned with AD and NC subjects as auxiliary domain to further improve the classification performance. Our method contains two key components: (1) the cross-domain kernel learning for transferring auxiliary domain knowledge, and (2) the adapted support vector machine (SVM) decision function construction for cross-domain and auxiliary domain knowledge fusion. Experimental results on the Alzheimer's Disease neuroimaging initiative (ADNI) database show that the proposed method can significantly improve the classification performance between MCI-C and MCI-NC, with aid of domain knowledge learned from AD and NC subjects.

Fig. 1

Distributions of AD, NC, MCI-C, and MCI-NC subjects with CSF features

Fig. 2

Flow chart of the proposed DTSVM classification method

Fig. 3

ROC curves of DTSVM, LapSVM, and SVM, using multi-modality and single-modality data, respectively

Fig. 4

Comparison of classification accuracy of DTSVM and LapSVM with respect to the use of different number of subjects in the auxiliary domain