Prediction of Mild Cognitive Impairment Conversion Using a Combination of Independent Component Analysis and the Cox Model

Mild cognitive impairment (MCI) represents a transitional stage from normal aging to Alzheimer's disease (AD) and corresponds to a higher risk of developing AD. Thus, it is necessary to explore and predict the onset of AD in MCI stage. In this study, we propose a combination of independent component analysis (ICA) and the multivariate Cox proportional hazards regression model to investigate promising risk factors associated with MCI conversion among 126 MCI converters and 108 MCI non-converters from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Using structural magnetic resonance imaging (MRI) and fluorodeoxyglucose positron emission tomography (FDG-PET) data, we extracted brain networks from AD and normal control groups via ICA and then constructed Cox models that included network-based neuroimaging factors for the MCI group. We carried out five separate Cox analyses and the two-modality neuroimaging Cox model identified three significant network-based risk factors with higher prediction performance (accuracy = 73.50%) than those in either single-modality model (accuracy = 68.80%). Additionally, the results of the comprehensive Cox model, including significant neuroimaging factors and clinical variables, demonstrated that MCI individuals with reduced gray matter volume in a temporal lobe-related network of structural MRI hazard ratio (HR) = 8.29E-05 (95% confidence interval (CI), 5.10E- 07 ~ 0.013), low glucose metabolism in the posterior default mode network based on FDG-PET HR = 0.066 (95% CI, 4.63E-03 ~ 0.928), positive apolipoprotein E ε4-status HR = 1. 988 (95% CI, 1.531 ~ 2.581), increased Alzheimer's Disease Assessment Scale-Cognitive Subscale scores HR = 1.100 (95% CI, 1.059 ~ 1.144) and Sum of Boxes of Clinical Dementia Rating scores HR = 1.622 (95% CI, 1.364 ~ 1.930) were more likely to convert to AD within 36 months after baselines. These significant risk factors in such comprehensive Cox model had the best prediction ability (accuracy = 84.62%, sensitivity = 86.51%, specificity = 82.41%) compared to either neuroimaging factors or clinical variables alone. These results suggested that a combination of ICA and Cox model analyses could be used successfully in survival analysis and provide a network-based perspective of MCI progression or AD-related studies.