Machine learning for missing data imputation in Alzheimer's research: predicting medial temporal lobe dynamic flexibility.

Alzheimer's disease (AD) pathology begins years before symptoms appear, and dynamic flexibility of the medial temporal lobe (MTL) may serve as an early functional biomarker. Using data from 656 older adults in the Rutgers Aging and Brain Health Alliance study, we evaluated whether cognitive, genetic, biochemical, and demographic predictors could estimate MTL dynamic flexibility, despite substantial missingness (1,866 missing values; 25.86%). Only 42 participants (6.40%) had complete data; therefore, we compared case deletion with five imputation strategies (MICE, GAIN, MissForest, MIWAE, ReMasker) and eight regression models, assessing prediction accuracy using repeated 5-fold cross-validation. Complete-case analysis yielded limited performance (average [Formula: see text], [Formula: see text]). After imputation, all methods improved accuracy, with MissForest paired with Bagging Trees or Random Forest achieving the lowest prediction error ([Formula: see text]). The greatest improvement in concordance occurred when GAIN was combined with Bagging Trees/Random Forest ([Formula: see text]), representing a  57% gain over the best complete-case model. A Scheirer-Ray-Hare ANOVA confirmed significant differences across imputation strategies ([Formula: see text]). Runtime analyses showed GAIN and MissForest to be both accurate and computationally efficient, while deep generative imputers were slower. These findings demonstrate that robust imputation is essential for maximizing data utility and predictive reliability in high-missingness neuroimaging studies and highlight the potential of ensemble tree models combined with advanced imputation techniques for estimating MTL dynamic flexibility in aging populations.