Project Summary/Abstract A plethora of neuroscience studies shows mounting evidence that neurodegenerative diseases manifest distinct network dysfunction patterns much earlier prior to the onset of clinical symptoms. Since the subject-specific longitudinal network changes are more relevant to the neuropathological process than topological patterns derived from cross-sectional data, recognizing the subtle and dynamic longitudinal network biomarkers from noisy network data is of great demand to enhance the sensitivity and specificity of computer-assisted diagnosis in neurodegenerative diseases. However, current popular statistical inference or machine learning approaches used for neuroimages (in a regular data structure such as grid and lattice) are not fully optimized for the learning task on brain network data which is often encoded in a high dimensional graph (an irregular and non-linear data structure). Such gross adaption is partially responsible for the lack of reliable biomarkers that can be used to predict cognitive decline in routine clinical practice. To address this challenge, we aim to (1) develop a novel GNN (graph neural network) based learning framework to hierarchically discover the multi-scale network biomarkers that can recognize the disease-relevant network alterations over time, and (2) examine the diagnostic power of the new network biomarkers derived from our GNN-based machine learning engine across neurodegenerative diseases such as Alzheimer’s disease, Parkinson’s disease, and frontotemporal dementia. The success of this project will allow us to integrate the novel GNN-based learning component into our current longitudinal network analysis toolbox and release the AI (artificial intelligence) based network analysis software to the neuroscience and neuroimaging community.