Background: Automated multi-label classification of chest radiographs is complicated by overlapping imaging findings, noisy report-derived labels, and changes in disease co-occurrence across clinical settings. Existing clinical-hierarchy methods can encode label relationships, but a rigid hierarchy may conflate categorical membership with empirical association and propagate annotation errors. Methods: We investigated Reliability-Aware Clinical Relation Learning (ReCaRL), a graph-assisted framework for multi-label chest radiograph classification.