Toward precision brain health: accurate prediction of a cognitive index trajectory using neuroimaging metrics

Abstract

The goal of precision brain health is to accurately predict individuals' longitudinal patterns of brain change. We trained a machine learning model to predict changes in a cognitive index of brain health from neurophysiologic metrics. A total of 48 participants (ages 21-65) completed a sensorimotor task during 2 functional magnetic resonance imaging sessions 6 mo apart. Hemodynamic response functions (HRFs) were parameterized using traditional (amplitude, dispersion, latency) and novel (curvature, canonicality) metrics, serving as inputs to a neural network model that predicted gain on indices of brain health (cognitive factor scores) for each participant. The optimal neural network model successfully predicted substantial gain on the cognitive index of brain health with 90% accuracy (determined by 5-fold cross-validation) from 3 HRF parameters: amplitude change, dispersion change, and similarity to a canonical HRF shape at baseline. For individuals with canonical baseline HRFs, substantial gain in the index is overwhelmingly predicted by decreases in HRF amplitude. For individuals with non-canonical baseline HRFs, substantial gain in the index is predicted by congruent changes in both HRF amplitude and dispersion. Our results illustrate that neuroimaging measures can track cognitive indices in healthy states, and that machine learning approaches using novel metrics take important steps toward precision brain health.

Keywords: BOLD-fMRI; hemodynamic response; longitudinal study; machine learning; sensorimotor task.