From Brain Networks to Logical Rules: Monotonic GNNs for Alzheimer’s Disease

Graph deep learning methods are ubiquitously applied in the domain of brain network analysis. However, informal post-hoc interpretability analysis limits the application of these models for proposing hypotheses for clinical validation. In this work, we investigate monotonic GNNs whose predictions admit formally sound logical explanations which could be used to identify candidate structural biomarkers associated with Alzheimer's Disease. We demonstrate competitive results against established deep learning baselines and address some of the scalability challenges of extracting logical rules for noisy neuroimaging datasets. We provide an initial proof-of-concept rule to motivate further improvements in the rule extraction algorithm on large neuroimaging connectomes. Our results highlight that formal faithfulness provides guarantees of model behaviour, but does not itself establish clinical validity of an extracted explanation. Analyses are performed on large-scale neuroimaging study data from ADNI and OASIS-3.