Calibration-Aware and Interpretable Graph Learning for Multi-Cohort Diffusion Connectome Brain-Age Modeling

Researchers have created a novel approach to derive brain-age models from diffusion MRI-based structural connectomes. This method could potentially serve as an imaging biomarker for accelerated brain aging, but its applicability across diverse populations is still unclear. The study examined four distinct aging and Alzheimer's disease cohorts: ADNI, Duke/UNC ADRC, HABS-HD, and AD-DECODE, involving 1,093 connectome sessions from 789 participants.

Cohort-specific graph neural networks were trained using cross-validation, considering five different combinations of imaging and multimodal features. Prediction accuracy varied significantly between cohorts, with the imaging-only graph neural network achieving mean absolute error ranging from 4.72 years in ADNI to 9.75 years in AD-DECODE.

The imaging-only graph neural network performed comparably to ridge, elastic-net, and gradient-boosted regression models based on matched vectorized connectome features, but it was not consistently superior.

A key finding was that the age-bias-corrected brain-age gap was most consistently linked to reduced microstructural integrity and structural-network organization across the cohorts. In longitudinal analyses, this gap demonstrated moderate-to-good preservation within individuals, particularly in ADNI and HABS-HD, as indicated by intraclass correlation coefficients of 0.67 and 0.81, respectively. Higher initial values also predicted subsequent deterioration in microstructural and network aspects in ADNI.

Through a multiscale SHAP analysis, researchers identified contributions from various sources, including global graph topology, regional imaging features, edge-derived summaries, and individual structural connections involving multiple brain regions like the thalamus, striatum, frontal, parietal, cerebellar, hippocampal, and entorhinal circuits.

However, external transferability was highly sensitive to changes between cohorts. After recalibrating target-cohort models to match the chronological age of the source cohorts, the median mean absolute error decreased significantly, while the median Pearson correlation remained relatively low (0.17).

The recalibration process was primarily used for diagnostic sensitivity analysis rather than for deploying external validation models. Overall, these findings suggest that calibration-aware diffusion-connectome brain age can serve as an interpretable imaging biomarker for structural brain aging and prospective microstructural and network vulnerability. However, it emphasizes the critical need for cohort-specific calibration before applying such models to external datasets.