Lung ultrasound (LUS) diagnosis requires integrating findings across multiple anatomical zones, often under incomplete acquisition. We propose ZAGNet, a zone-aware graph neural network for patient-level LUS classification. Frame-level detections of consolidation and pleural effusion are temporally linked into tracklets and represented as nodes with spatial and confidence features. Nodes are connected using an anatomical adjacency map, and message passing aggregates inter-zone context. A global node produces patient-level predictions, enabling flexible handling of variable and missing zones. Evaluated on a multi-center dataset of 714 subjects (20,256 video loops), ZAGNet achieved AUC 0.803 for consolidation and 0.893 for pleural effusion, outperforming max (0.677/0.804) and mean pooling (0.674/0.815). Results show that graph-based aggregation captures inter-zone dependencies and supports robust patient-level LUS assessment.
