GraphPL: GNN-Based Modality Imputation for Patchwork Learning

GraphPL, an innovative technique, merges graph neural networks with patchwork learning to address the issue of missing modalities in distributed multi-modal learning. Existing methods typically presume that clients possess full modality data, a scenario that is frequently impractical. By incorporating all available modalities, GraphPL maintains its effectiveness even with noisy data, setting a new standard in benchmark evaluations. Evaluations conducted on a real-world distributed electronic health record dataset reveal that GraphPL excels in extracting robust downstream features for applications such as disease prediction.