Improving behavioral prediction from individual functional connectivity through transfer learning

1813 

Abstract Submission 

Alexandre Le Bris¹

¹Inria, Palaiseau, Essonne

Alexandre Le Bris  
Inria
Palaiseau, Essonne

Resting-state functional magnetic resonance imaging (rfMRI) enables the exploration of the functional brain organization and its representation via large-scale networks [1]. Summary measures, such as functional connectivity (FC) or networks' spatial topography, present interindividual differences with applications in detecting cognitive disorders or predicting behavioral traits [2,3,4]. However, the limited number of subjects in clinical data can significantly disrupt these estimations and then the accuracy of downstream tasks. We show here that this difficulty can be partially overcome with transfer learning (TL) coupled with Bayesian modeling.

The dataset comprises 1,000 subjects from the S1200 release of the HCP dataset [5], each going with 2 rfMRI scans acquired on consecutive days. For a given subject and run, the input data represent the functional connectivity of each vertex of the cortical mesh with reference regions provided by the DiFuMo atlas [6].

We considered that each functional connectome is the combination of two individual-level components: a probabilistic parcellation of the cortex into a finite number of networks and a set of network-specific FC profiles. This is depicted through a Bayesian model, where these components are random variables (RVs) to be inferred. Notably, individual variables are regularized by population priors (Fig. 1A).

To estimate the latent RVs, we employed PAVI [7], a variational inference technique leveraging probabilistic graphical models and deep learning techniques like normalizing flows [8]. This approach allows simultaneous estimation of cortical topography and FC at the individual level. Following the extraction of these features, inference performance is evaluated through the regression of 58 behavioral scores provided by the HCP dataset.

In a first experiment, we assessed the model's learning capacity by varying the number of subjects used to infer the features and then examining the evolution of mean accuracy when predicting behavioral scores.

Secondly, we divided the 1,000 subjects into a large dataset of 750 individuals (the source dataset) and a smaller one ranging from 50 to 200 subjects (out-of-sample or target dataset). We pre-trained a model with the source dataset and used the inferred features to transfer part of the learning to a new model trained on the out-of-sample dataset. We tested two TL approaches. The first, a simple warm start, involves transferring pre-trained parameters in PAVI neural networks to the second model. The second approach, in addition to the warm start, utilizes population priors trained on the source dataset and, as such, is denoted by Bayesian TL (Fig. 1B).

Evaluating the model's learning capacity reveals a logarithmic growth in prediction accuracy as the number of subjects in the dataset grows. Up to 1,000 subjects, the slope indicates that full learning capacity has not been reached (Fig. 2A). Further subjects are required to enhance prediction performance, justifying the TL approach.

The second experiment illustrates the efficiency of TL in inferring individual summary measures, such as subject-specific FC profiles, from a limited dataset. With Bayesian TL and 200 previously unseen subjects, we achieve comparable performance in behavioral prediction as directly inferring with twice as many subjects (Fig. 2A). However, this outcome, averaged across 58 behavioral scores, exhibits considerable variation among the considered behavioral measures. Some even undergo negative TL (Fig. 2B). The reasons for such variability (e.g. summary measures not capturing the underlying neural cause) are yet to be fully understood.

These initial results demonstrate the advantages of TL in deriving individually customized representations of functional brain organization with a limited dataset. However, further confirmation is needed when applying this approach to heterogeneous data from diverse datasets.

Bayesian Modeling

fMRI Connectivity and Network Modeling ¹

Methods Development ²

Segmentation and Parcellation

Computational Neuroscience

FUNCTIONAL MRI

Machine Learning

Modeling

Segmentation