A fully inductive inference protocol for population GNNs in single-subject brain disorder diagnosis

Population graph-based Graph Neural Networks (GNNs) have demonstrated superior performance in brain disease diagnosis by modeling inter-subject relationships. However, most existing approaches rely on a transductive setting that achieves high performance on known subjects but suffers from significant performance degradation when applied to unseen subjects. While a few inductive population graph models have been proposed, they struggle with single-subject inference, either due to a reliance on test batches for graph construction or limited generalization capabilities for individual unseen nodes. In this paper, we propose a fully inductive inference protocol in population graphs designed for single-subject diagnosis. Our approach constructs a population graph exclusively with training nodes and dynamically establishes connections between a single unseen test subject and the training graph during the inference phase based on imaging and phenotypic similarities. We conducted extensive experiments on multiple neuroimaging datasets (ABIDE I, ABIDE II, and ADHD-200) to evaluate the proposed pipeline. The results demonstrate that our method outperforms both state-of-the-art transductive models and existing inductive baselines under a fully inductive evaluation protocol. Furthermore, our analysis reveals that single-subject inference tends to maximize diagnostic performance within our experimental settings by reducing potential interference between test subjects. Importantly, our approach obviates the prohibitive retraining bottleneck typically required by transductive models, thereby providing an operational advantage for deployment and facilitating efficient, real-time single-subject inference workflows. The source code is available at https://github.com/98jaemin/single_subject_popgnn.