BrainOSM: Outlier screening for multi-view functional brain network analysis

Identifying biomarkers for mental diseases is vital for understanding their underlying mechanisms, facilitating early diagnosis, and enabling more personalized treatment strategies. In this study, we focus on diagnosing autism spectrum disorder (ASD) and alzheimer's disease (AD) by analyzing functional brain networks (FBNs), which are represented as graphs capturing the functional connectivity patterns of the brain. The primary challenges in modeling FBNs for this disorder stem from two key issues: (i) the heterogeneity among graphs, and (ii) the disease-unrelated information within graphs.

We introduce a two-stage framework, BrainOSM, which combines outlier screening in datasets with a multi-view graph pooling module for enhanced graph classification. Specifically, the first stage employs progressive uncertainty-based outlier screening to reduce the interference of inter-graph heterogeneity. The second stage integrates multi-graph pooling, multi-view learning, and prior subnetwork regularization to refine graph structures, effectively tackling the challenge of disease-unrelated information within graphs.

To validate the effectiveness of our method, we assess its performance on two public datasets: the Autism Brain Imaging Data Exchange (ABIDE) dataset and the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. On the ABIDE dataset, BrainOSM achieved an average accuracy of 70.23% and an AUC of 70.42%, corresponding to improvements of 8.55% and 7.74% over the traditional GCN method. On the ADNI dataset, it reached an average accuracy of 82.29% and an AUC of 83.23%, showing gains of 8.97% and 11.78%, respectively. Our code is publicly available at https://github.com/guoguiliang111/BrainOSM.

Our extensive experiments confirm the generalizability and the effectiveness of BrainOSM for mental disease classification. Visual analyses further demonstrate that the model effectively identifies subnetworks associated with mental diseases, highlighting its potential for clinical interpretation. Moreover, our findings indicate that outlier screening plays a crucial role in improving classification accuracy when dealing with heterogeneous datasets.