基于多模态相关向量回归机的老年痴呆症临床变量预测

Predicting clinical variables in Alzheimer's disease based on multimodal relevance vector regression

老年痴呆症(Alzheimer's disease,AD)的临床变量值和多模态特征都是对其内在致病病理的外在反映.本文提出一种多模态相关向量回归机,通过对多模态特征的学习来预测临床变量值.首先采用核方法将多模态数据融合成一个混合核矩阵,然后使用相关向量回归机对临床变量简易精神状态检查(mini mental state examination,MMSE)和老年痴呆症评定量表(Alzheimer's disease assessment scale,ADAS-Cog)建立回归模型,最后用相关系数和平方根均方误差来验证算法的性能.在标准数据集ADNI上的实验结果表明,本文提出的多模态方法的预测性能优于单模态方法.

Recently, effective and accurate diagnosis disease stage of Alzheimer＇s disease （AD） or mild cognitive impairment （MCI） has attracted more and more attention. Numerous studies have demonstrated that clinical variables and multimodal features of AD are external reflections of the intrinsic disease pathology. This paper proposes a multimodal regression method for estimating disease stage and predicting clinical progression from three modalities of biomarkers, i. e., magnetic resonance imaging （MRI）, fluoro-deoxy-glucose-positron emission tomography （ FDG-PET）, and cerebrospinal fluid （CSF） biomarkers. Specifically, our multimodal regression framework includes three key steps： firstly, we use the specific application tool to orginal MRI and FDG-PET images data from the 202 Alzheimer＇ s disease neuroimaging initiative（ADNI） subjects. For each preproeessed original MR or FDG PET image, 93 regions of interest （ROIs） are labeled by an atlas warping algorithm. And then, for each MR or FDG-PET image, 93 volumetric features are extracted from the 93 ROIs. Therefore, for each subject, the last features come from 93 features from the MRI image, another 93 features from the PET image, and 3 features from the CSF biomarkers which original values are directly used as features. Secondly, muhi-modal data （MRI, PET, and CSF） are fused into a mixed kernel matrix through the kernel method, and then regression model is constructed by the relevant vector machine regression （RVR） for the clinical variables including mini mental state examination（MMSE） and Alzheimer＇ s disease assessment scale（ADAS-Cog）. Finally, our multimodal regression method compared with single-modal approaches and bimodal approaches. Moreover, the performance of our regression model is validated by the correlation coefficient （CORR） and the square root of the mean square error （RMSE）. These regression experiment scheme are tested on the ADNI dataset by 10-fold cross-validation. Experimental results on the ADNI database show that the prediction performance of our multi-modal RVR approach is superior to the corresponding single-modal approaches and bimodal approaches.