A Hybrid Spatial Statistical Learning Framework for Individualized Probability Estimation: Application to Multisite Autism Neuroimaging
Year: 2026
Type: preprint
Abstract: Autism spectrum disorder is a heterogeneous neurodevelopmental condition whose functional brain organization varies across individuals and imaging centers. Resting-state functional connectivity provides an opportunity to investigate this variation, but its analysis is complicated by high dimensionality, strong dependence among connections, and substantial differences in acquisition protocols and participant populations across sites. Existing approaches commonly rely on either detailed connectivity features or lower-dimensional network summaries and are often evaluated primarily through discrimination. A model may therefore rank participants reasonably well while producing probability estimates that are unreliable when transported to a new imaging environment. We introduce a Hybrid Spatial Statistical Learning Framework for individualized probability estimation in multisite autism neuroimaging. The framework treats edge-level connectivity, graph-based network organization, and participant characteristics as complementary representations of a common probability target. Detailed connectivity retains fine-scale discriminatory information, while broader representations are used to stabilize the probability scale under site heterogeneity. All model development and preprocessing are performed within complete site-held-out validation, providing a direct assessment of transportability to unseen acquisition environments. Simulation studies show that the hybrid preserves the discrimination of the strongest fine-scale model while improving probability accuracy as between-site heterogeneity increases. In an analysis of 860 participants from 20 Autism Brain Imaging Data Exchange sites, the hybrid retained the edge model's ROC AUC, increasing it only from 0.646 to 0.647, while reducing the Brier score from 0.256 to 0.243 and the log loss from 0.733 to 0.684. These represent reductions of 4.9% and 6.6%, respectively. The hybrid reduced Brier-score error at 18 of the 20 held-out sites. These findings demonstrate that multiscale integration can improve the reliability and transportability of individualized autism probabilities. More broadly, the framework provides an interpretable approach for probability estimation from complex, spatially dependent biomedical data collected across heterogeneous institutions.
Source: arXiv (Cornell University)
Institutions: St. Edward's University, Rice University
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7204108531