Effects of uncertainty contributions on the examination of Alzheimer’s disease via MEG data and graph neural networks

The multidimensional nature of the information encoded in magnetoencephalography (MEG) data demands for decision-making tools based on graph topology. In this context, possible sources of uncertainty pertinent to the instrumentation, the inferential blocks, and to the interaction of the test subject with the measurement system have to be taken into account to derive adequate examinations. We adopt techniques in the emerging field of geometric deep learning (GDL), particularly the graph neural networks (GNNs), and propose a novel framework for the examination of familiarity with Alzheimer’s disease (AD) in MEG recordings. The core idea is that two GNNs, trained on MEG data of reference subjects of each category, i.e., with or without familiarity, respectively, can be used for extracting graph-based representations and temporal data predictions for the evaluation of the brain-network activity of test subjects during the execution of a memory task. We show that quantifying graph topology and similarity of test recordings with temporal predictions of one or the other GNN can help the decision-making process. To test the robustness of the approach, three main uncertainty contributions including instrumental, inferential, and interaction uncertainty are modeled and propagated throughout the system via a Monte Carlo approach. We analyze their effects on the learned-graph representations and on the final examination. On a set of 50 subjects, an accuracy of 0.76 (0.03) was obtained with leave-one-patient out cross-validation in the scenario that combines all the three uncertainty scenarios, proving the effectiveness of the proposed methods. Performance improvements of more than 10% have been achieved by combining multiple reference GNNs with an ensemble approach.