How to shape connectivity in task-evoked functional MRI data: An examination

1560 

Abstract Submission 

Alice Giubergia^1,2, Sara Mascheretti³, Valentina Lampis¹, Tommaso Ciceri^1,2, Martina Villa⁴, Chiara Andreola⁵, Filippo Arrigoni⁶, Alessandra Bertoldo², Denis Peruzzo¹

¹Scientific Institute IRCCS Eugenio Medea, Bosisio Parini (LC), Italy, ²University of Padova, Padova, Italy, ³University of Pavia, Pavia, Italy, ⁴University of Connecticut, Storrs, CT, ⁵Université Paris Cité, Paris, Paris, ⁶V. Buzzi Children’s Hospital, Milan, Italy

Alice Giubergia  
Scientific Institute IRCCS Eugenio Medea|University of Padova
Bosisio Parini (LC), Italy|Padova, Italy

Sara Mascheretti  
University of Pavia
Pavia, Italy

Valentina Lampis  
Scientific Institute IRCCS Eugenio Medea
Bosisio Parini (LC), Italy

Tommaso Ciceri  
Scientific Institute IRCCS Eugenio Medea|University of Padova
Bosisio Parini (LC), Italy|Padova, Italy

Martina Villa  
University of Connecticut
Storrs, CT

Chiara Andreola  
Université Paris Cité
Paris, Paris

Filippo Arrigoni  
V. Buzzi Children’s Hospital
Milan, Italy

Alessandra Bertoldo  
University of Padova
Padova, Italy

Denis Peruzzo  
Scientific Institute IRCCS Eugenio Medea
Bosisio Parini (LC), Italy

Functional connectomics, which maps functional associations among brain regions, typically relies on resting-state fMRI (Biswal 2010). However, due to practical constraints, many studies lack sufficient resting-state data (Elliott 2019) and resort to inferring a "pseudo-resting" state through task-modulated connectivity (Bhandari 2020). Task connectivity allows for examining broader brain region involvement beyond task activation (Di 2019). This study aims to assess the feasibility of inferring "pseudo-resting" state connectivity from task data and explore its impact on subsequent behavioral trait analysis.

Seventy-seven subjects (age 9-18 years, M/F: 49/28), 39 Typical Readers – TR and 38 with Developmental Dyslexia – DD) underwent two visual tasks, Sinusoidal Gratings (SG) and Coherent Motion (CM), using fMRI. BOLD time-series were processed with FreeSurfer to obtain task (i.e., without task regression) and "pseudo-resting" (i.e., after task regression) conditions for both tasks across 200 cortical and 18 deep Grey Matter (dGM) regions of interest (ROIs). Various task regression setups were tested varying the number of derivatives in the GLM's HRF function, and connectomes were derived using ROI-wise Pearson correlation of each subject's time series, resulting in 218×218 matrices. Spurious connections were addressed through thresholding (He 2010) and multiple task-classification experiments were conducted using different Cross-Validated (CV) machine learning algorithms. Figure 1 reports SG Vs. CM classification metrics. After that, we set the sparsity level (50%) and the derivatives for task regression (d=0), and performed a classification experiment with the aim of investigating the impact of task regression in the characterization of DD and TR. In each fold of each experiment accuracy, AUC score, and the connections selected for the classification were saved to dig in the discriminative process. Connections selected in at least 5 folds out of 6 were identified as discriminative and compared between the 4 SVM experiments (i.e., task SG, task CM, "pseudo-resting" SG and "pseudo-resting" CM). All connections were grouped by the belonging macro-ROIs as defined in the Schaefer atlas to get an overview of the areas devoted to the discrimination of DD and TR.

All task classification experiments provided significant classification performances, with a classification accuracy and AUC larger than 50%. Task regression did remove task-related content from fMRI signals, but the stimulus could still be inferred from derived connectomes irrespective of preprocessing. SVM experiments demonstrated successful classification between Typical Readers (TR) and children with Developmental Dyslexia (DD) in both task and "pseudo-resting" conditions (p<0.001). Circular plots of Figure 2 display the preprocessing-modulated connections. No connections were commonly selected between SG and CM tasks, both in task and "pseudo-resting" states (intra-processing and inter-task). However, over 50% of connections were commonly selected when comparing task and "pseudo-resting" conditions for SG and CM (intra-task and inter-processing). Macro-ROI analysis highlighted the involvement of regions known to be associated with DD, while task-related differences aligned with distinct networks elicited by each task.

Task fMRI data preprocessing was examined to understand the impact of task information on connectomics. Successful discrimination in "pseudo-resting" conditions suggests that connectomes retain task information even after GLM regression. This implies that a task-free "pseudo-resting" state cannot be reliably inferred from task fMRI. Discriminative connections depend on the task type, influencing how classifiers differentiate between groups. Despite task-related differences, signal preprocessing does not significantly affect classification rule inference, indicating similar evaluation of significant features in both tasks.

Neurodevelopmental/ Early Life (eg. ADHD, autism)

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural) ¹

fMRI Connectivity and Network Modeling

Data analysis

FUNCTIONAL MRI

MRI

Pediatric Disorders