Transferable individual brain regional atrophy progression prediction model using structural connectome

Abstract

AbstractBackgroundBrain microstructure, as measured by diffusion MR imaging (dMRI), is an early biomarker of Alzheimer’s disease (AD) and has been previously demonstrated to predict whole brain atrophy and cognition. The network‐based neurodegeneration hypothesis posits that the onset of AD occurs within key vulnerable regions, whose network connectivity guides the spread of atrophy into new regions. Graph convolutional network (GCN) models can mimic this spread through its convolution process. Here we present a GCN‐based model to predict future individual nodal atrophy and disease progression across the AD spectrum using structural connectome.MethodWe analyzed 371 participants with baseline dMRI scans (154 cognitively normal (CN), 161 mild cognitive impairment (MCI), 56 AD) and longitudinal T1 scans (2.7±1.2 years) from the ADNI study. To predict annual nodal atrophy rates (126 nodes) from baseline structural connectivity (SC), we constructed a two‐layer GCN followed by multilayer perceptron (GCN‐MLP). Five‐fold nested cross‐validation and Optuna were used for hyperparameter tuning. We applied the trained GCN‐MLP model to predict future disease progression via transfer learning. GNNExplainer was used to identify the predictive SC features for atrophy or cognition. The GCN‐MLP model was validated in another independent dataset from Singapore (MACC) with 226 participants (61 CN, 112 MCI, 53 AD, follow‐up period of T1 scans (3.1±1.3 years)) using transfer learning.ResultThe proposed GCN‐MLP model predicted the annual rate of regional atrophy at the individual level in the main (R = 0.5407±0.0480; MAE = 0.0013±0.0001; EV = 0.2930±0.0516) and validation (R = 0.5323±0.0564; MAE = 0.0009±0.0001; EV = 0.2808±0.0559) datasets (Table 1; Figure 1). Graph edge SC of the target network contributed most to the atrophy prediction of its own network (Figure 2A&C). Nodal feature SC in hippocampus, striatum, and temporal network were major predictors of atrophy progression across all 10 networks (Figure 2B&D).Moreover, the trained atrophy prediction model predicted changes of CDR‐Sum‐of‐Boxes (main: R = 0.8035±0.0492; validation: R = 0.5555±0.0404), ADAS13 (main: R = 0.7475±0.0725), MMSE (main: R = 0.7124±0.0509; validation: R = 0.6377±0.0616), and MOCA (validation: R = 0.6764±0.0314).ConclusionWe developed an accurate and interpretable individualized longitudinal nodal atrophy and disease progression prediction model in the AD spectrum by leveraging the comprehensive properties of whole‐brain structural connectome. The approach was validated in both Caucasian and Asian datasets, highlighting its generalizability and interpretability.