In the network approach to psychopathology, disorders can be represented as communities of co-occurring symptoms, with comorbidity attributed partly to bridge symptoms that link them. Identifying these bridge symptoms is challenging because existing approaches struggle to disentangle a symptom's role within its community from its role in connecting communities. Moreover, the network is itself uncertain, as edges are inferred rather than observed and their presence varies from one sample to the next. 

To disentangle these roles, we group edges rather than nodes into communities via the line graph transformation. A node whose edges span multiple communities emerges as a bridge, quantified by a continuous bridge score for each node. To propagate network uncertainty, bridge scores are averaged across posterior samples from a Bayesian Ordinal Markov Random Field with a Stochastic Block Model prior. In simulation, we derive bridge scores from curvature-based and minimum-cut partitions of the network and its line graph, with Louvain as benchmark, varying community separability, bridge membership strength, and cluster balance. 

Curvature on the line graph achieved the lowest Brier score in nearly all conditions, with the gap narrowing under low separability. Louvain's gain was similar but smaller, while minimum-cut occasionally performed worse on the line graph. All variants degraded under strong cluster imbalance, because nodes have more potential connections to the larger community. In a reanalysis of a 371-participant depression-anxiety dataset, line-graph bridge scores correlated near zero or negatively with traditional bridge centrality, indicating they are not merely a refinement, but capture a distinct structural quantity.