Unsupervised Dynamic Graph Multi-Model Representation Learning for Temporal Patterns Discovery: Uncovering Parkinson’s Disease Stages Using Cerebrospinal Fluid Longitudinal Profiles

ICLR 2026 Conference Submission25570 Authors

Keywords: Learning Representation, Dynamic Graphs, Parkinson’s Disease, deep learning, unsupervised learning

TL;DR: We created a multi-model graph learning method that integrates node representations across graph snapshots, capturing temporal trajectories and spatial context.

Abstract: Existing dynamic graph learning methods typically encode node features at each time step by leveraging local (spatial/ structural) and/or short-range temporal dependencies. In contrast, we propose a novel multi-model framework that generates a representation for each node at every graph snapshot, where each representation encodes the node’s temporal trajectory across the full sequence while preserving its spatial context within that specific time step. When clustered, these representations reveal meaningful temporal pattern groups in longitudinal datasets. This approach was evaluated in the context of Parkinson’s disease (PD), a degenerative disorder that progresses in distinct clinical stages. To demonstrate this, we structured six years of longitudinal cerebrospinal fluid (CSF) records from 24 patients with PD into age-based graphs, where clinical visit records corresponding to patients with the same age are hosted on the graph for that age. In these graphs, nodes represent individual patients indexed by unique identifiers and are enriched with CSF peptide abundance features. Edges are established between patient nodes based on the similarity of their peptide expression patterns. For each patient's node, a one-layer Graph Convolutional Network (GCN) was employed to encode inter-patient relationships within each age-specific graph. The resulting spatial representations across all time points for each node were then fed into a sequential model to learn a unified spatial-temporal features representation for every patient. To represent patient’s features at each age, the unified embedding was combined with the age-specific GCN representation through a fusion block - composed of linear transformations, nonlinear activations, and normalization layers - producing rich, locally informed spatial embeddings that are further enhanced by the global context of inter- and intra-related node patterns. K-means++ clustering of the multi-model representations identified four distinct disease stages, supported by strong cluster validity metrics (Davies-Bouldin Index = 0.169, Calinski-Harabasz Index = 1264.24). When statistically analysed, Kruskal-Wallis test revealed significant differences in motor scores (UPDRS_2 and UPDRS_3; p < 0.05) across clusters, with Dunn’s test further identifying which clusters differed significantly. Unlike the motor scores, where most patient profiles apparently clustered into two groups, the non-motor scores (UPDRS_1) were distributed across three clusters but did not show significant differences (p = 0.11). The learned embeddings revealed well-separated clinical motor profiles, outperforming PCA, autoencoders, GCN, T-GCN, and GC-LSTM representations in capturing clinically relevant dimensions of disease severity. With further optimization and validation, this framework could aid in staging and understanding neurodegenerative diseases and generalizes to other longitudinal pattern discovery tasks.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 25570