Graph Artificial Intelligence in Medicine

Abstract

In clinical artificial intelligence (AI), graph representation learning, mainly through graph neural networks and graph transformer architectures, stands out for its capability to capture intricate relationships and structures within clinical datasets. With diverse data-from patient records to imaging-graph AI models process data holistically by viewing modalities and entities within them as nodes interconnected by their relationships. Graph AI facilitates model transfer across clinical tasks, enabling models to generalize across patient populations without additional parameters and with minimal to no retraining. However, the importance of human-centered design and model interpretability in clinical decision-making cannot be overstated. Since graph AI models capture information through localized neural transformations defined on relational datasets, they offer both an opportunity and a challenge in elucidating model rationale. Knowledge graphs can enhance interpretability by aligning model-driven insights with medical knowledge. Emerging graph AI models integrate diverse data modalities through pretraining, facilitate interactive feedback loops, and foster human-AI collaboration, paving the way toward clinically meaningful predictions.

Keywords: artificial intelligence; graph neural networks; graph transformers; health care; human-centered AI; knowledge graphs; medicine; multimodal learning; transfer learning.

Figure 1:

Variants of neural networks for graphs, including temporal graphs, hypergraphs, and subgraphs. a) shows a message-passing scheme by which messages are propagated along edges for each node. b) shows an example of a graph transformer, where the graph is expanded into a sequence with positional encodings. c) shows temporal graphs sampled over times $t,t+1$; these methods often use variants of a) or b) to learn representations at each step. d) shows a hypergraph, where hyperedges (rounded shapes) can encompass more than two nodes. e) shows a subgraph neural network, which learns representations for $S_i$, $S_j$ even when these subgraphs are disconnected.

Figure 2:

Strategies for incorporating inductive biases and enabling transfer learning on knowledge graphs, including a) biologically informed neural network architectures, b) supplementing patient data with biomedical ontologies, and c) fine-tuning large pre-trained models across a broad range of clinical tasks.

Figure 3:

Deep learning on a) multimodal clinical datasets including omics-data, medical imaging, EHR, and more. Data modalities can be fused through b) early, c) intermediate, and d) late integration strategies. Elements of this figure were created using BioRender.

Figure 4:

Mathematical and user-centered explainability for clinical AI tools affect various stakeholders across data scientists, researchers, clinicians, patients, and administrators.