Quantifying Multi-site Heterogeneity in Tractography-Based Regression of SRS Cognition in Autism Spectrum Disorder

Author First name, Last name, Institution

• Mohamed Khudri, J.B. Speed School of Engineering
• Mostafa Abdelrahim, J.B. Speed School of Engineering
• Moumen El-Melegy, J.B. Speed School of Engineering
• Ali Mahmoud, J.B. Speed School of Engineering
• Asem Ali, J.B. Speed School of Engineering
• Ahmed Shalaby, UT Southwestern Medical Center
• Mohammed A. Ghazal, Abu Dhabi University
• Fatma Taher, Zayed University, Abu Dhabi Campus
• Sohail Contractor, University of Louisville School of Medicine
• Gregory N. Barnes, University of Louisville School of Medicine
• Ayman El-Baz, J.B. Speed School of Engineering

Document Type

Conference Proceeding

Source of Publication

Lecture Notes in Computer Science

Publication Date

8-2-2026

Abstract

Diffusion MRI tractography provides a powerful view into how long-range white-matter pathways support distributed brain communication and cognition. Building on this capability, we introduce a tractography-based regression framework for autism spectrum disorder (ASD) that predicts continuous SRS cognition scores while explicitly quantifying the impact of multi-site acquisition heterogeneity. Diffusion MRI from 226 ABIDE-II participants were processed; 219 subjects with cognition scores across five sites (BNI_1, NYU_1, NYU_2, SDSU_1, TCD_1) were included in all analyses. Whole-brain structural connectomes were generated using an MRtrix3-centered pipeline and vectorized into connectivity features. Dimensionality was reduced using Spearman and mutual-information screening followed by Ridge-based recursive feature elimination. Multiple regression models including linear, kernel-based, tree ensemble, Bayesian, and deep/foundation models were evaluated with Optuna under nested cross-validation, and pooled multi-site learning was systematically compared with site-specific modeling under site-preserving imbalance handling (SMOGN and sample reweighting). Pooled models delivered stable intermediate performance, with TabPFN achieving the best pooled results under no augmentation (MAE =3.22±1.22, RMSE =6.37±1.46, R²=0.88±0.05, Spearman ρ=0.93±0.03). Site-specific performance was strongly site-dependent: NYU cohorts achieved the lowest errors (median MAE ≈1.3–3.5), whereas higher-variability sites showed substantially larger errors (TCD_1 MAE ≈14–16). Paired Wilcoxon tests on test MAE confirmed significant regime differences with large effects, favoring site-specific learning at NYU_1 (p=0.002, FDR ≤0.005, Cliff’s δ=-1.00) and pooled learning at TCD_1 (FDR ≤0.005, Cliff’s δ=1.00). These results provide a systematic benchmark of pooled versus site-specific tractography regression for continuous cognition modeling.

DOI Link

10.1007/978-3-032-31930-2_8

ISBN

[9783032319296]

ISSN

0302-9743

Publisher

Springer Nature Switzerland

Volume

16825 LNCS

First Page

110

Last Page

125

Disciplines

Computer Sciences | Medicine and Health Sciences

Keywords

Autism, Diffusion MRI, ML, Multi-site Learning, Site Heterogeneity, Tractography

Scopus ID

105047066799

Recommended Citation

Khudri, Mohamed; Abdelrahim, Mostafa; El-Melegy, Moumen; Mahmoud, Ali; Ali, Asem; Shalaby, Ahmed; Ghazal, Mohammed A.; Taher, Fatma; Contractor, Sohail; Barnes, Gregory N.; and El-Baz, Ayman, "Quantifying Multi-site Heterogeneity in Tractography-Based Regression of SRS Cognition in Autism Spectrum Disorder" (2026). All Works. 8039.
https://zuscholars.zu.ac.ae/works/8039

Indexed in Scopus

yes

Open Access

no