Artificial intelligence may be exceptionally good at recognizing patterns in medical data, yet that strength can become a weakness when algorithms leave the hospitals and populations in which they were trained. A system that performs impressively in one institution can lose accuracy in another, where patient demographics, clinical practices, equipment and data-recording habits differ. Worse, it may preserve discriminatory patterns hidden in historical records. A new Perspective in Nature Biomedical Engineering examines causal graph neural networks as a possible route toward healthcare AI that is more robust, interpretable and adaptable across settings.
The central problem is that conventional machine-learning systems often learn statistical associations rather than the biological or clinical mechanisms that generate those associations. For example, an algorithm might connect a diagnosis with a particular scanner, hospital department or demographic characteristic because those features happened to correlate in its training data. If the surrounding conditions change, the connection can fail. Causal approaches attempt to identify relationships that remain stable when the environment changes, focusing on what influences an outcome rather than merely what predicts it.
Causal graph neural networks combine two powerful ideas. Graph neural networks represent complex systems as networks of connected entities, allowing information to move between related nodes. In healthcare, nodes might represent genes, proteins, brain regions, symptoms, medications, physiological measurements or patients, while edges describe interactions or dependencies. Structural causal models add a formal description of how variables are generated, often through equations that distinguish direct causes from observed effects and unmeasured influences. Together, these tools can model biomedical systems as interconnected mechanisms rather than as flat tables of features.
The Perspective, authored by M. Mesinovic, M. Buhlan and T. Zhu, reviews several routes for building such models. One is disentangled causal representation learning, in which an AI system tries to separate different underlying factors that are mixed together in raw medical data. A brain scan, for instance, may reflect anatomy, disease state, age, motion during imaging and scanner-specific artifacts at the same time. A disentangled model aims to represent these influences separately, making it easier to determine which factors are clinically meaningful and which are accidental features of data collection.
Another key capability is interventional prediction. Standard prediction asks what is likely to happen when the system is observed as it is. Causal inference asks a different question: what would happen if a variable were actively changed? In a healthcare setting, that could mean estimating the effect of altering a medication, modifying a treatment plan or targeting a biological pathway. Graph-based causal models can encode these hypothetical changes by intervening on selected nodes and propagating their consequences through the network. This is more demanding than correlation-based prediction because it requires assumptions about the structure of the system.
Counterfactual reasoning goes a step further by asking what would have happened to a particular patient under an alternative scenario. A model might estimate how a patient’s physiological trajectory could have differed if a drug had not been prescribed, or whether a different intervention might have reduced risk. These calculations are not direct observations; they are simulations based on the model’s structure, data and assumptions. The authors therefore emphasize that counterfactual outputs should not automatically be treated as proven clinical facts.
The potential applications span several areas of medicine. In psychiatry, causal graph neural networks could help connect symptoms, brain-network activity and treatment responses, where diagnoses are often heterogeneous and biological mechanisms remain difficult to isolate. In cancer research, they could integrate genomic, transcriptomic and other multi-omics measurements to identify regulatory relationships associated with tumor subtypes. In continuous monitoring, graph-based models could link streams such as heart rate, blood pressure, oxygen saturation and activity to detect evolving physiological states and forecast clinical deterioration.
The same framework could also support more informed drug recommendations by estimating how interventions might affect individual patients. This vision leads to the idea of a patient-specific causal digital twin: a computational representation designed to simulate disease progression and test possible interventions in silico before they are applied in the clinic. Such systems could eventually help researchers explore treatment strategies, identify likely responders and reduce unnecessary experimentation on patients. However, a digital twin would only be as reliable as the causal relationships, measurements and validation procedures used to construct it.
Major obstacles remain before these methods can move from promising research tools to dependable clinical systems. Causal discovery is computationally expensive, particularly when graphs contain thousands of biological variables and data arrive continuously. Real-time deployment may therefore be difficult. Clinical data are also incomplete, noisy and shaped by treatment decisions, meaning that the records used to learn causal relationships may themselves reflect bias. A model can produce a sophisticated-looking graph without having established that its edges represent genuine causal mechanisms.
Validation must consequently go beyond ordinary cross-validation, which tests whether a model predicts well on held-out data drawn from a similar distribution. Stronger evaluation may require testing across hospitals, populations, instruments and time periods, as well as comparing predictions with results from randomized or carefully designed observational studies. The authors warn against “causal-washing,” in which causal language is applied to methods that have not earned causal credibility. They propose a tiered framework distinguishing causally inspired architectures, models that incorporate explicit causal assumptions, and causally validated discoveries supported by independent evidence. For causal AI to deliver on its promise, the field will need scalable algorithms, multimodal integration, transparent assumptions and regulatory standards calibrated to the strength of each claim.
Subject of Research: Causal graph neural networks and causal artificial intelligence methods for healthcare.
Article Title: Causal graph neural networks for healthcare.
Article References: Mesinovic, M., Buhlan, M. & Zhu, T. Causal graph neural networks for healthcare. Nature Biomedical Engineering (2026). https://doi.org/10.1038/s41551-026-01742-3
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41551-026-01742-3
Keywords: causal artificial intelligence, graph neural networks, healthcare AI, causal inference, structural causal models, counterfactual reasoning, digital twins, precision medicine, multimodal biomedical data, clinical machine learning
