Predicting Cognitive Abilities with Graph Signal Processing Metrics
Poster No:
1412
Submission Type:
Abstract Submission
Authors:
Mikkel Schöttner1, Thomas Bolton2, Jagruti Patel3, Patric Hagmann2
Institutions:
1University of Lausanne, Lausanne, Vaud, 2Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL), Lausanne, Switzerland, 3CHUV, Zurich, Not Specified
First Author:
Co-Author(s):
Thomas Bolton
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland
Patric Hagmann
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland
Introduction:
Predicting behavior from neuroimaging data comes with the prospects of selecting an optimal treatment, detecting mental illness at an early stage, or forecasting cognitive development (Wu et al., 2023). As a notable example, cognitive features have been predicted from functional connectivity (FC) through kernel ridge regression (KRR) (He et al., 2020). However, KRR does not allow direct inference about which brain networks drive the prediction, as it uses the whole feature vector for training.
Graph signal processing (GSP) is a multimodal framework that expresses functional brain time series in terms of a basis of connectome harmonics that contrast structural brain features across spatial frequencies (Huang et al., 2018) . Advanced GSP measures like the structural decoupling index (SDI) (Preti & Van De Ville, 2019) enable efficient fingerprinting (Griffa et al., 2022), and might prove useful for behavioral prediction as well.
Here, we explore how GSP metrics, combined with an elastic-net regression scheme, compare to the FC/KRR approach in predicting a factor reflective of cognition. We also pinpoint optimal parameter choices for this task.
Graph signal processing (GSP) is a multimodal framework that expresses functional brain time series in terms of a basis of connectome harmonics that contrast structural brain features across spatial frequencies (Huang et al., 2018) . Advanced GSP measures like the structural decoupling index (SDI) (Preti & Van De Ville, 2019) enable efficient fingerprinting (Griffa et al., 2022), and might prove useful for behavioral prediction as well.
Here, we explore how GSP metrics, combined with an elastic-net regression scheme, compare to the FC/KRR approach in predicting a factor reflective of cognition. We also pinpoint optimal parameter choices for this task.
Methods:
We used 847 subjects from the Human Connectome Project (HCP) dataset (Van Essen et al., 2012). As a predictive target, we chose one of four factors that optimally summarize the behavioral variables from the HCP dataset (Schöttner et al., 2022): cognition.
For each subject, using T1-weighted and diffusion-weighted imaging data, deterministic fiber tracking was performed with Connectome Mapper 3 (Tourbier et al., 2022), and a structural connectivity (SC) matrix was created using the Lausanne 2018 atlas (Cammoun et al., 2012) at scale 3 (R=274 regions) with normalized fiber density as edge weight. Resting-state functional magnetic resonance imaging (fMRI) data were detrended, regressed for six motion parameters and their derivatives, and parcellated using the same atlas. Finally, the regional time series were z-scored.
Two GSP features were computed: the SDI, which measures how much the functional signals are constrained by the underlying structural connectivity (higher values = less constrained) (Preti & Van De Ville, 2019), and the power spectral density (PSD) of the fMRI signals across spatial frequencies. As graph shift operator (GSO, the quantity that connectome harmonics are derived from), we compared the normalized Laplacian and the modularity matrix (Petrovic et al., 2020). For the SDI, a cutoff frequency separating low and high frequencies must be chosen. This can involve (1) a half-half split (i.e., cutting at R/2), (2) choosing the frequency that splits cumulative PSD in half for each time point (individual cumulative sum), (3) using the mode over time points (cumulative sum), or (4) following (2) on the mean PSD spectrum over time (mean cumulative sum).
For prediction, an elastic-net regression was evaluated using nested cross-validation with 10 repeated random splits in the inner and outer loops, placing members of the same family in the same split. As a measure of prediction accuracy, we considered the coefficient of determination (R²). The same evaluation scheme was used for FC/KRR as a baseline.
For each subject, using T1-weighted and diffusion-weighted imaging data, deterministic fiber tracking was performed with Connectome Mapper 3 (Tourbier et al., 2022), and a structural connectivity (SC) matrix was created using the Lausanne 2018 atlas (Cammoun et al., 2012) at scale 3 (R=274 regions) with normalized fiber density as edge weight. Resting-state functional magnetic resonance imaging (fMRI) data were detrended, regressed for six motion parameters and their derivatives, and parcellated using the same atlas. Finally, the regional time series were z-scored.
Two GSP features were computed: the SDI, which measures how much the functional signals are constrained by the underlying structural connectivity (higher values = less constrained) (Preti & Van De Ville, 2019), and the power spectral density (PSD) of the fMRI signals across spatial frequencies. As graph shift operator (GSO, the quantity that connectome harmonics are derived from), we compared the normalized Laplacian and the modularity matrix (Petrovic et al., 2020). For the SDI, a cutoff frequency separating low and high frequencies must be chosen. This can involve (1) a half-half split (i.e., cutting at R/2), (2) choosing the frequency that splits cumulative PSD in half for each time point (individual cumulative sum), (3) using the mode over time points (cumulative sum), or (4) following (2) on the mean PSD spectrum over time (mean cumulative sum).
For prediction, an elastic-net regression was evaluated using nested cross-validation with 10 repeated random splits in the inner and outer loops, placing members of the same family in the same split. As a measure of prediction accuracy, we considered the coefficient of determination (R²). The same evaluation scheme was used for FC/KRR as a baseline.
Results:
Both PSD and SDI yielded a median predictive performance close to the baseline with the right parameters (Fig. 1A). In terms of GSO, the normalized Laplacian seemed to work better for PSD and the modularity matrix for the SDI, for which splitting the signal by half was also the optimal choice. Fig. 1B shows the absolute values of the beta coefficients mapped to the brain surface for the best-performing model using the SDI. The largest coefficients were in the left and right superior parietal cortex, and the right inferior parietal cortex.
Conclusions:
GSP-derived metrics enable similar predictive performance as FC to predict a factor reflective of cognition, with the added benefit of allowing the direct investigation of what regional features drive the prediction in the case of the SDI.
Higher Cognitive Functions:
Executive Function, Cognitive Control and Decision Making 2
Reasoning and Problem Solving
Higher Cognitive Functions Other
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Connectivity (eg. functional, effective, structural)
Keywords:
Cognition
Data analysis
FUNCTIONAL MRI
Machine Learning
MRI
STRUCTURAL MRI
1|2Indicates the priority used for review
Provide references using author date format
Griffa, A. (2022). Brain structure-function coupling provides signatures for task decoding and individual fingerprinting. NeuroImage, 250, 118970.
He, T. (2020). Deep neural networks and kernel regression achieve comparable accuracies for functional connectivity prediction of behavior and demographics. NeuroImage, 206, 116276
Huang, W., Bolton, T. A. W. (2018). A Graph Signal Processing Perspective on Functional Brain Imaging. Proceedings of the IEEE, 106(5), 868–885.
Petrovic, M. (2020). Community-Aware Graph Signal Processing: Modularity Defines New Ways of Processing Graph Signals. IEEE Signal Processing Magazine, 37(6), 150–159.
Preti, M. G. (2019). Decoupling of brain function from structure reveals regional behavioral specialization in humans. Nature Communications, 10(1), Article 1.
Schöttner, M. (2022, June 19). Domains of Behavior and Psychopathology Derived via Factor Analysis. Annual Meeting of the Organization for Human Brain Mapping. Annual Meeting of the Organization for Human Brain Mapping, Glasgow, United Kingdom.
Tourbier, S. (2022). Connectome Mapper 3: A Flexible and Open-Source Pipeline Software for Multiscale Multimodal Human Connectome Mapping. Journal of Open Source Software, 7(74), 4248.
Van Essen, D. C. (2012). The Human Connectome Project: A data acquisition perspective. NeuroImage, 62(4), 2222–2231.
Wu, J. (2023). The challenges and prospects of brain-based prediction of behaviour. Nature Human Behaviour, 7(8), Article 8.