MMGL

MMGL predicts disease outcomes by integrating multi-modal medical data with graph-based representation learning to model inter-sample relationships.

Key Features:

• End-to-End Framework: Performs end-to-end learning from raw multi-modal inputs to disease prediction without requiring manual graph construction.
• Modality-Aware Representation Learning: Aggregates features across modalities such as demographic information and clinical measurements to capture inter-modality correlations and complementarities.
• Adaptive Graph Learning: Learns latent graph structures adaptively and jointly optimizes the graph with the prediction model to uncover intrinsic connections among samples.
• Inductive Learning Capability: Supports inductive inference to handle unseen patient data at prediction time.

Scientific Applications:

• Disease Prediction: Applied to disease prediction tasks across biomedical fields by integrating heterogeneous modalities into a cohesive graph representation.
• Diagnostic Modeling: Enhances diagnostic modeling by capturing complex inter-modality correlations for more reliable predictions.
• Personalized Medicine: Supports personalized medicine and improved diagnostic accuracy by jointly integrating modalities and adaptively learning patient-sample relationships.

Methodology:

Uses modality-aware representation learning and adaptive graph learning, where latent graph structures are dynamically constructed and refined and jointly optimized with the prediction model, with support for inductive inference.