Neuroimaging-based Individualized Prediction of Cognition and Behavior for Mental Disorders and Health: Methods and Promises

Abstract

The neuroimaging community has witnessed a paradigm shift in biomarker discovery from using traditional univariate brain mapping approaches to multivariate predictive models, allowing the field to move toward a translational neuroscience era. Regression-based multivariate models (hereafter "predictive modeling") provide a powerful and widely used approach to predict human behavior with neuroimaging features. These studies maintain a focus on decoding individual differences in a continuously behavioral phenotype from neuroimaging data, opening up an exciting opportunity to describe the human brain at the single-subject level. In this survey, we provide an overview of recent studies that utilize machine learning approaches to identify neuroimaging predictors over the past decade. We first review regression-based approaches and highlight connectome-based predictive modeling, which has grown in popularity in recent years. Next, we systematically describe recent representative studies using these tools in the context of cognitive function, symptom severity, personality traits, and emotion processing. Finally, we highlight a few challenges related to combining multimodal data, longitudinal prediction, external validations, and the employment of deep learning methods that have emerged from our review of the existing literature, as well as present some promising and challenging future directions.

Figure 1.. Visual summary of studies using regression-based machine learning approaches to predict continuous variables.

(A) There is an obviously increasing trend in the number of papers published each year since 2010. (B) The overall prediction accuracy against the corresponding sample size used in the studies. (C) The type of behaviors of interest that are used as the target measures among all surveyed studies. (D) Cognitive metrics adopted in our surveyed studies. (E) Distribution of prediction accuracy for healthy subjects and patients with brain disorders shown as boxplot plot (left) and kernel density (right). (F) Distribution of prediction accuracy from studies using multimodal or unimodal data.

Figure 2.. Summary of regression approaches used in our surveyed papers.

Multi-task approaches jointly predict multiple clinical variables in a unified framework, while single-task methods only predict one type of cognitive score at one time. Most of the surveyed papers used linear models to reveal brain-behavior relationships. Connectome-based predictive modeling (CPM) is a recently developed data-driven approach that combines simple linear regression and feature selection together to predict individual differences in traits and behavior from connectivity data, and has been successfully employed for the prediction of multiple human behaviors.

Figure 3.. Overview of the connectome-based predictive modeling (CPM) framework and its applications in behavior prediction.

(A) Overview of general analysis strategy for CPM procedure. Specifically, CPM is performed first by calculating the relevance of each connectivity feature to behavioral measure of interest across subjects and retaining the most significantly correlated ones under a predefined threshold. And then, a single aggregate metric named ‘network strength’ is computed by summing strength of the retained connectivity features. Afterwards, the summary statistics and behavioral scores are submitted to a simple linear regression model. By placing the procedure within a cross-validation framework, accurate estimations of behavioral scores can be obtained. (B) Beaty et al. accomplished robust prediction of individual creative ability using FCs acquired from 163 participants engaged in a classic divergent thinking task, and assessed the generalizability in three external cohorts (n=39, 54, and 405). (C) Rosenberg et al. accomplished robust prediction of sustained attention using CPM in a sample of 25 healthy subjects, and the identified neuromarkers were generalized to predict a clinical measure for patients with ADHD (validation cohort 1, n=113) and individuals’ go response rate in a stop-signal task (validation cohort 2, n=83). (Adapted from ref. (48), and (20, 36)).

Figure 4.

Using multimodal data to predict cognition promotes the biomarker identification. (A) Jiang et al. achieved an improved prediction performance of intelligence scores by integrating FCs and cortical thickness. More importantly, the study suggested that these two types of neuroimaging features provided unique evidence of the neurobiological correlates of intelligence from distinct perspective. Specifically, prediction with cortical thickness explored more gender difference in the lateralization of predictive brain regions, while prediction with FCs detected more gender difference in the specification of contributing functional networks. (B) Sui et al. predicted the cognitive composite scores for subjects from three independent cohorts based on neuromarkers derived from a supervised multimodal fusion approach (Adapted from ref. (86), and (87)).