A hybrid algorithm of machine learning and optimization to identify Super‐Agers using rsfMRI data

Abstract

AbstractBackgroundAging and cognitive decline are often associated with each other. However, some adults called Super‐Agers have superior cognitive performance relative to age, with a comparable cognitive performance against healthy middle‐aged individuals. This study employs resting‐state functional Magnetic Resonance Imaging (rsfMRI) and demographic information to distinguish Super‐Agers from Cognitive Decliners.MethodFirst, Principal Component Analysis (PCA) is applied to a set of four cognition tests which were recorded for two‐time points, (baseline visit in 2006‐2010) and (follow‐up visit starting in 2014) to create the initial latent cognitive trajectories groups. Then, a hybrid algorithm of machine learning and optimization is developed to predict the latent groups using orthogonal neural networks. The proposed algorithm utilizes Bayesian Optimization to optimize the labeling procedure and a logistic regression model to classify participants.ResultWe compared the performance of the proposed algorithm against two groups of baseline models: 1) four models, each using only one of the cognition tests at a time for labeling; 2) two models, each with a restricted set of features (one using only demographics information, and the other one using rsfMRI data). The proposed algorithm outperformed the baseline models, achieving an AUC of 88% when distinguishing between Super‐Agers and Cognitive Decliners.ConclusionThe proposed algorithm is capable of distinguishing cognitive trajectories with high success. The combination of demographics and rsfMRI datasets improved the overall performance compared to individual datasets.