Predicting Risk of Conversion to Clinically Diagnosed Alzheimer’s Disease from Baseline sMRI Using Supervised Functional Principal Component Analysis

Accurate prediction of the risk of progression to clinically diagnosed Alzheimer’s Disease (AD) in cognitively normal (CN) individuals and those with mild cognitive impairment (MCI) is essential for early intervention and prognosis. Structural magnetic resonance imaging (sMRI) provides rich information on brain morphology, but its high dimensionality poses challenges for robust risk prediction, especially with limited sample sizes and censored outcomes. In this study, we extend a prediction-oriented framework based on supervised functional principal component analysis (sFPCA) to extract low-dimensional components from baseline sMRI data in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort. sFPCA extracts components that balance dominant imaging variation against their association with the survival outcome, thereby reducing the dimensionality of the sMRI data while retaining relevant imaging information. The resulting components were evaluated using both Cox proportional hazards (Cox PHs) and random survival forest (RSF) models, providing complementary log-linear and nonparametric survival-prediction frameworks. Across ten retained development/test partitions, mean test IAUC was 0.7733 for sFPCA-Cox and 0.7294 for sFPCA-RSF, with corresponding Integrated Brier Scores (IBSs) of 0.08730 and 0.09232, respectively. The matched uFPCA models achieved mean IAUC values of 0.7719 and 0.7357 under Cox PH and RSF, respectively, indicating that the incremental predictive contribution of survival supervision was modest and depended on the downstream survival model. While the primary objective is risk prediction, additional analysis results based on low-, middle- and high-risk groups are used to facilitate interpretation of the learned imaging components. The identified anatomical spatial patterns involve regions such as the hippocampus, amygdala, and prefrontal and temporal cortices, supporting their relevance to AD-related neurodegeneration. These findings demonstrate the feasibility of sMRI-based prediction of conversion to clinically diagnosed AD and suggest that survival supervision may provide modest improvements under some modeling configurations.