Causal Graph Neural Networks for Healthcare

TL;DR

This paper argues that healthcare AI must move beyond associative pattern recognition to causal reasoning to survive distribution shifts, reduce discrimination, and provide mechanistic interpretability. It synthesizes the integration of structural causal models with graph neural networks, detailing methods for disentangling causal signals, interventional prediction, counterfactual generation, robustness, and fairness. Across diagnoses, prognoses, treatments, and real-time monitoring, the authors showcase how causal GNNs can reveal true biological mechanisms, enable patient-specific simulations, and guide mechanism-based therapies, culminating in the aspirational framework of Causal Digital Twins. They also outline substantial barriers—computational demands, validation challenges, regulatory gaps, and risks of causal-washing—and propose a tiered evidentiary framework to guide future research and clinical translation.

Abstract

Healthcare artificial intelligence systems routinely fail when deployed across institutions, with documented performance drops and perpetuation of discriminatory patterns embedded in historical data. This brittleness stems, in part, from learning statistical associations rather than causal mechanisms. Causal graph neural networks address this triple crisis of distribution shift, discrimination, and inscrutability by combining graph-based representations of biomedical data with causal inference principles to learn invariant mechanisms rather than spurious correlations. This Review examines methodological foundations spanning structural causal models, disentangled causal representation learning, and techniques for interventional prediction and counterfactual reasoning on graphs. We analyse applications demonstrating clinical value across psychiatric diagnosis through brain network analysis, cancer subtyping via multi-omics causal integration, continuous physiological monitoring with mechanistic interpretation, and drug recommendation correcting prescription bias. These advances establish foundations for patient-specific Causal Digital Twins, enabling in silico clinical experimentation, with integration of large language models for hypothesis generation and causal graph neural networks for mechanistic validation. Substantial barriers remain, including computational requirements precluding real-time deployment, validation challenges demanding multi-modal evidence triangulation beyond cross-validation, and risks of causal-washing where methods employ causal terminology without rigorous evidentiary support. We propose tiered frameworks distinguishing causally-inspired architectures from causally-validated discoveries and identify critical research priorities making causal rather than purely associational claims.

Figures (3)

• Figure 1: Clinical failure modes of traditional machine learning and solutions through causal graph neural networks. A, Clinical failure modes demonstrate fundamental limitations of correlation-based models in healthcare deployment. A.1, Vulnerability to distribution shift: a diabetic retinopathy screening system achieves 94% accuracy at the training institution (Hospital A) but performance collapses to 73% at deployment sites (Hospital B) due to spurious correlations with site-specific imaging protocols and patient demographics rather than causal disease mechanisms. Feature distributions show that spurious features (red, site-specific) undergo a distributional shift, while causal features (blue, disease markers) remain invariant across environments. A.2, Lack of mechanistic insight: traditional models operate as black boxes, learning associations between age, treatment, and comorbidity without identifying the underlying causal pathways, precluding clinical interpretability and mechanistic validation. Question marks indicate uncertain causal relationships that cannot be disentangled from observational data alone. A.3, Counterfactual blindness: Associational models trained on observational data (Treatment A, 87% five-year survival) cannot simulate alternative treatment scenarios (Treatments B and C) marked with prohibition symbols, as they lack the structural causal knowledge required for interventional reasoning, which is essential to personalised treatment selection. B, Causal GNN solutions address these limitations through principled causal inference. B.1, Causal invariance learning: multi-environment optimisation identifies stable causal relationships (blue nodes with shield protection) whilst suppressing environment-specific spurious correlations (red nodes, faded), ensuring robust generalisation under distribution shift through invariant risk minimisation. B.2, Causal digital twin for counterfactual generation: integration of multi-modal patient data (shown as concentric rings) constructs patient-specific structural causal models encoding disease mechanisms. Application of the do-operator enables simulation of unobserved treatment scenarios, generating counterfactual predictions (Treatment B: 73%, Treatment C: 95% five-year survival) through interventional inference rather than requiring empirical observations, facilitating treatment optimisation for individual patients.
• Figure 2: Conceptual framework illustrating theoretical advantages of causal over associational inference across Pearl's causal hierarchy. A, Theoretical dissociation between associational and interventional performance predicted by causal hierarchy theory. The "causal gap" quantifies expected performance loss when models trained on observational data are applied to interventional queries, reflecting the fundamental limitation that associational patterns may not persist under intervention. B, Idealised robustness to simulated unmeasured confounding. Conceptual curves illustrate the theoretical expectation that traditional methods degrade more steeply as unmeasured confounder strength increases, while causal architectures that explicitly model confounding structure are expected to maintain robustness. Confounding strength parameterised as maximum risk ratio for unmeasured confounder associations with treatment and outcome following sensitivity analysis frameworks VanderWeele_2017_Sensitivity. Shaded regions represent idealised uncertainty bands illustrating expected variance patterns. Causally-Inspired GNN (Tier 1) denotes architectures employing causal concepts such as invariance regularisation to improve robustness without explicit causal claims. Causally-Validated GNN (Tier 3) denotes methods whose causal claims are corroborated through multi-modal evidence triangulation. See Section 6.3 for formal tier definitions.
• Figure 3: Canonical causal graph neural network architectures illustrated for conceptual brain connectivity analysis. While the underlying methods were developed on synthetic benchmark datasets, we illustrate their application to functional neuroimaging to demonstrate translation potential for clinical contexts. A, The DisC (Debiasing via Disentangled Causal Substructure) framework Fan_2022_Debiasing achieves causal disentanglement through a dual-encoder architecture. An edge mask generator partitions input graphs into causal and spurious subgraphs, with separate GNN modules encoding each into disentangled representations. The loss comprises two components: $\mathcal{L}_{CE}$, a weighted cross-entropy loss that up-weights samples where the bias GNN struggles, directing the causal encoder toward genuinely predictive features; and $\mathcal{L}_{GCE}$, a generalised cross-entropy loss that amplifies gradients on easily-classified samples, directing the bias encoder toward environment-specific shortcuts. Causal features ($Z_{C.1}$--$Z_{C.4}$) produce stable predictions across environments while spurious features ($Z_{S.1}$--$Z_{S.3}$) are explicitly identified and discarded. B, The iVGAE (interventional Variational Graph Auto-Encoder) framework Zecevic_2021_Relating implements Pearl's do-operator through differentiable neural layers. The interventional encoder maps observed graphs to a causal latent space $Z$, while the interventional decoder incorporates do-layers that simulate graph surgery. The output $P(\mathbf{D}|\textit{do}(\mathbf{X}))$ represents the interventional distribution of downstream variables $\mathbf{D}$ given intervention $\textit{do}(\mathbf{X}=x)$, enabling prediction of effects that persist under manipulation rather than spurious correlations reflecting confounding. Applied to brain networks, such architectures could distinguish genuine neural communication pathways from global arousal-induced correlations, identifying intervention targets for neuromodulation or pharmacotherapy.