Assessing Frequency-Dependent Behavior Predictability via Coherence-based Fingerprinting

1734 

Abstract Submission 

Wenjun Bai¹, Okito Yamashita¹, Junichiro Yoshimoto¹

¹CBI, ATR, Kyoto, Kyoto

Wenjun Bai  
CBI, ATR
Kyoto, Kyoto

Okito Yamashita  
CBI, ATR
Kyoto, Kyoto

Junichiro Yoshimoto, Dr.  
CBI, ATR
Kyoto, Kyoto

Leveraging vast resting-state functional MRI (rsfMRI) data and machine learning, individual-level rsfMRI-based brain connectivity, especially the functional connectome, serves as reliable brain features in capturing individual behavioral differences (Finn et al. 2015, Beaty et al. 2018, Nostro et al. 2018). Despite its limited frequency bandwidth (0.01-0.20Hz (Niazy et al. 2011, Yuen et al. 2019)), rsfMRI signals show a diverse spectral distribution in the human cortex (Fox&Raichle2007), prompting the question of whether these spectral differences influence intrapersonal behavior prediction. To answer this question, we propose the coherence-based predictive modeling (CoPM), using spectrally rich functional coherence features to explore this spectral-behavior relationship. Our application of CoPM to Human Connectome Project rsfMRI data (Van Essen et al. 2013) reveals that the behavior predictability of functional coherence features is frequency-dependent, highlighting a strong spectral-behavior relationship.

The proposed coherence-based predictive model (CoPM) aims to evaluate the predictive performance of frequency-specialized coherence features across various behavioral items and domains. The methodology of CoPM is structured into a three-stage pipeline as follows: (A) brain-wide coherence feature extraction (including Table 1 embedded in Figure 1), (B) frequency-dependent coherence profiling (including Table 2 embedded in Figure 1), and (C) predictive model construction. A graphical representation of this pipeline is illustrated in Figure 1. (For a complete demonstration of the proposed CoPM, please refer to the OHBM2024_complete.pdf file at https://github.com/LeonBai/Rhythmic_Cortex.)

Upon applying the proposed CoPM to the HCP rsfMRI data, our initial observation highlighted the discriminative capability of functional coherence features in identifying intrapersonal behavioral differences, regardless of their frequency. Notably, within the predicted behavior domains, psychological well-being emerged as the consistent domain that was successfully predicted across all six coherence profiles. This suggests that brain signatures related to well-being may encompass a broad spectral range. Similarly, the personality behavior domain also showed predictions across a wide frequency bandwidth, expect for the low-frequency profile-6. In contrast, successful predictions for the emotion domain were predominantly associated with the median-frequency coherence features (as shown in Figure 2(a)).

The predictive performance varied among the six profiles, with high-frequency profile-2 showing the least success, and median-frequency profiles (profile-3, 4, and 5) demonstrating broader and more accurate predictions. This frequency-dependent prediction was further evident when the similarity between profile-wise rhythmic (e.g., frequency) characteristics and the prediction patterns were observed in terms of a strong positive correlation (Pearson's r = 0.606,p < 0.001; Figure 2(b)).

The widespread use of rsfMRI recordings and machine learning advancements have enabled individual-level cortico-cortical connections, like the functional connectome, to serve as crucial brain signatures for identifying intrapersonal behavioral differences. As cortical regions show varied spectral distributions, exploring their spectral bias in behavior prediction is gaining interest. Our study introduces coherence-based predictive modeling (CoPM), a novel approach leveraging spectrally rich coherence features to explore their frequency-dependent behavior predictability. Applied to HCP rsfMRI data, CoPM effectively discerns individual behavioral differences using frequency-dependent coherence features. It demonstrates that behavioral predictability is frequency-dependent, offering valuable insights for future CoPM applications in diverse neuroimaging contexts, such as exploiting the use of functional coherence features in computational psychiatry.

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural)

fMRI Connectivity and Network Modeling ¹

Workflows

Informatics Other

ADULTS

Computational Neuroscience

Data analysis

FUNCTIONAL MRI

Informatics

Machine Learning