Published: 2026-02-06 23:41
Intracranial aneurysms, often silent until they rupture, pose a significant challenge in neurovascular medicine. The unpredictable nature of their rupture can lead to devastating consequences, including subarachnoid haemorrhage, stroke, and death.
Clinicians face the complex task of identifying which unruptured aneurysms are most likely to rupture, a decision that heavily influences whether to pursue invasive treatment or watchful waiting.
A new development published in npj Digital Medicine introduces an artificial intelligence (AI) model designed to predict intracranial aneurysm hemodynamics in real time. This novel approach, utilising a physics-constrained graph neural network (GNN), could offer a rapid, non-invasive method to assess the intricate blood flow patterns within aneurysms, potentially transforming risk stratification and treatment planning for patients in the UK and beyond.
The Clinical Challenge of Intracranial Aneurysms
An intracranial aneurysm is a weak, bulging spot on the wall of a brain artery. While many remain asymptomatic throughout a person’s life, their rupture is a medical emergency with high morbidity and mortality.
The annual incidence of subarachnoid haemorrhage in the UK is estimated to be around 6-8 per 100,000 people, with a significant proportion caused by ruptured aneurysms.
The decision to treat an unruptured aneurysm is a delicate balance. Intervention carries its own risks, including stroke, haemorrhage, and neurological deficits.
Conversely, leaving a high-risk aneurysm untreated can result in catastrophic rupture. Current clinical guidelines, such as those from NICE, often consider factors like aneurysm size, location, patient age, and previous subarachnoid haemorrhage history in family members.
However, these factors alone do not always provide a complete picture of rupture risk.
Aneurysm morphology and, critically, the hemodynamics – the forces exerted by blood flow on the aneurysm wall – are increasingly recognised as crucial determinants of rupture risk. Abnormal blood flow patterns can lead to wall degeneration, inflammation, and ultimately, rupture.
Understanding these forces is paramount for informed clinical decisions.
Hemodynamics: A Key to Understanding Rupture Risk
Blood flow within an aneurysm is complex and highly individualised, influenced by the aneurysm’s unique shape, size, and the parent artery’s characteristics. Key hemodynamic parameters thought to be associated with aneurysm rupture include:
- Wall Shear Stress (WSS): The frictional force exerted by blood flow on the vessel wall. Both very low and very high WSS have been implicated in aneurysm growth and rupture. Low WSS can lead to chronic inflammation and endothelial dysfunction, while high WSS can cause direct mechanical stress.
- Oscillatory Shear Index (OSI): A measure of the directional change of WSS over the cardiac cycle. High OSI indicates highly disturbed and oscillating flow, which is often associated with aneurysm instability.
- Pressure Distribution: Localised areas of high pressure within the aneurysm sac can contribute to wall stress.
- Flow Patterns: Complex flow patterns, such as impingement zones where blood directly hits a specific part of the aneurysm wall, are also considered high-risk indicators.
Accurately assessing these parameters for each patient has traditionally been a significant hurdle. The gold standard for detailed hemodynamic analysis is Computational Fluid Dynamics (CFD).
Limitations of Traditional Computational Fluid Dynamics (CFD)
CFD involves creating a detailed 3D model of the aneurysm and surrounding vasculature from patient imaging data (e.g., CT angiography, MRA). Sophisticated mathematical equations are then used to simulate blood flow within this model. While highly accurate, CFD has several practical limitations in a busy clinical setting:
- Computational Intensity: Running CFD simulations requires significant computational power and time, often taking hours or even days to complete for a single aneurysm.
- Expertise Required: Interpreting CFD results and setting up simulations demands specialised engineering and fluid dynamics expertise, which is not routinely available in most clinical departments.
- Cost: The software and hardware required for advanced CFD analysis can be expensive.
- Lack of Real-Time Feedback: The time delay makes it impractical for immediate clinical decision-making or intraoperative guidance.
These limitations mean that despite its potential, CFD is not widely used in routine clinical practice for aneurysm risk assessment. This gap highlights the need for a faster, more accessible method to provide clinicians with crucial hemodynamic insights.
The Promise of AI: Physics-Constrained Graph Neural Networks
The new research addresses these challenges by leveraging the power of artificial intelligence, specifically a physics-constrained graph neural network (GNN). This innovative AI model aims to overcome the computational bottlenecks of traditional CFD, offering real-time hemodynamic predictions.
What is a Graph Neural Network (GNN)?
Graph neural networks are a class of neural networks designed to operate on data structured as graphs. In the context of intracranial aneurysms, the complex network of blood vessels and the aneurysm itself can be represented as a graph, where points on the vessel surface are ‘nodes’ and their connections are ‘edges’.
GNNs are particularly adept at learning relationships and patterns within such interconnected data structures, making them well-suited for analysing complex anatomical geometries and fluid dynamics.
The ‘Physics-Constrained’ Advantage
A crucial aspect of this new model is its “physics-constrained” nature. Traditional machine learning models often learn patterns purely from data, which can sometimes lead to predictions that are mathematically sound but physically implausible, especially when encountering novel or unusual data.
By incorporating fundamental laws of fluid dynamics (such as conservation of mass and momentum, described by Navier-Stokes equations) directly into the GNN’s learning process, the model is guided to produce physically consistent and accurate results.
This constraint helps the AI model to:
- Improve Accuracy: Ensures predictions adhere to known physical principles, enhancing reliability.
- Enhance Generalisability: Makes the model more robust when applied to a wider range of aneurysm geometries and patient-specific conditions, even those not extensively represented in the training data.
- Reduce Data Dependency: Less reliant on vast amounts of training data compared to purely data-driven models, as it already ‘knows’ some fundamental rules.
How the AI Model Works in Practice
The workflow for this AI model begins with standard patient imaging data, typically from CT angiography (CTA) or magnetic resonance angiography (MRA). These scans provide the detailed 3D anatomical information of the intracranial vasculature, including the aneurysm’s precise geometry.
- Image Segmentation and Mesh Generation: The 3D imaging data is processed to segment the aneurysm and its parent vessels. A computational mesh is then generated, discretising the vascular surface into a network of points (nodes) and connections (edges), forming the ‘graph’ structure for the GNN.
- Input to the GNN: This patient-specific anatomical graph, along with boundary conditions (e.g., estimated blood flow at the inlet of the parent artery), is fed into the trained physics-constrained GNN.
- Real-Time Prediction: The GNN rapidly processes this input. Instead of performing iterative, time-consuming simulations like CFD, the GNN leverages its learned understanding of fluid dynamics and geometry to directly predict the hemodynamic parameters across the aneurysm surface.
- Output Visualisation: The output includes detailed maps of wall shear stress, pressure, and flow velocity distributions, which can be visualised in real-time or near real-time.
The key differentiator is speed. Where traditional CFD might take hours, this AI model can potentially deliver results in minutes or even seconds, making it a viable tool for immediate clinical application.
Potential Clinical Applications for UK Neurovascular Teams
The ability to obtain rapid, accurate hemodynamic assessments could have a profound impact on several aspects of neurovascular care within the NHS.
Enhanced Risk Stratification
For patients presenting with unruptured intracranial aneurysms, the decision to treat or observe is critical. Current risk assessment tools are often insufficient to capture the full complexity of an individual aneurysm’s rupture potential. By providing detailed, patient-specific hemodynamic data, the AI model could:
- Identify High-Risk Aneurysms: Pinpoint aneurysms with adverse hemodynamic profiles (e.g., very low WSS, high OSI, strong impingement zones) that might warrant more aggressive management, even if they are small.
- Reassure on Low-Risk Aneurysms: Help clinicians confidently recommend watchful waiting for aneurysms with favourable hemodynamic characteristics, reducing unnecessary interventions and their associated risks.
- Improve Patient Counselling: Provide more objective and detailed information to patients and their families regarding their individual rupture risk, facilitating shared decision-making.
Optimised Treatment Planning
Once a decision to treat is made, selecting the appropriate intervention – surgical clipping or endovascular coiling – is crucial. The choice depends on various factors, including aneurysm morphology, location, and patient comorbidities. Hemodynamic insights can further refine this choice:
- Guiding Endovascular Coiling: For complex aneurysms, understanding flow patterns can help interventional neuroradiologists plan the optimal coiling strategy, predict the effectiveness of flow diverters, or identify areas prone to incomplete occlusion.
- Assessing Surgical Clipping Suitability: Surgeons could use the information to anticipate flow changes post-clipping and ensure complete exclusion of the aneurysm from circulation while preserving parent vessel flow.
- Personalised Device Selection: Potentially aid in selecting the most appropriate flow diverter or stent design based on pre-treatment hemodynamic predictions.
Post-Treatment Monitoring and Follow-Up
After treatment, patients require ongoing surveillance to monitor for recurrence or changes in residual aneurysms. The AI model could assist in:
- Early Detection of Recurrence: Rapidly assess changes in hemodynamics in treated aneurysms, potentially identifying early signs of recanalisation or growth that might require further intervention.
- Evaluating Treatment Efficacy: Quantify the hemodynamic changes achieved by an intervention, providing objective measures of treatment success.
Training and Education
Neurovascular training is highly specialised. The ability to simulate and visualise complex hemodynamics in real-time could be an invaluable educational tool for:
- Trainee Neurosurgeons and Neuroradiologists: Allowing them to explore the impact of different aneurysm morphologies and hypothetical interventions on blood flow without patient risk.
- Multidisciplinary Team (MDT) Discussions: Facilitating clearer communication and understanding among neurosurgeons, neuroradiologists, neurologists, and other specialists during complex case reviews.
Advantages for UK Healthcare
The adoption of such an AI model could bring several advantages to the UK healthcare system:
- Efficiency Gains: Faster analysis means quicker decision-making, potentially reducing patient waiting times for treatment plans and improving throughput in neurovascular clinics.
- Cost-Effectiveness: While initial investment in AI infrastructure might be required, the long-term benefits of reduced unnecessary interventions, improved outcomes, and more efficient use of specialist time could lead to cost savings.
- Improved Patient Outcomes: More accurate risk assessment and tailored treatment strategies are expected to lead to fewer aneurysm ruptures and better post-treatment results.
- Accessibility: If the computational demands are significantly lower than traditional CFD, the technology could become more accessible to a wider range of hospitals, not just large tertiary centres.
- Standardisation: Provides a more objective and standardised method for hemodynamic assessment, reducing inter-observer variability.
Limitations and Future Directions
While this research presents a significant leap forward, it is crucial to acknowledge the inherent limitations and the extensive work still required before such a model can be routinely deployed in UK clinical practice.
Rigorous Clinical Validation
The most critical next step is extensive prospective clinical validation. This involves testing the AI model on large, diverse cohorts of patients across multiple centres to confirm its accuracy, reliability, and clinical utility.
This validation must demonstrate that the AI’s predictions correlate strongly with actual patient outcomes and provide superior or equivalent performance to current clinical decision-making processes.
Generalisability Across Diverse Aneurysm Types
Aneurysms exhibit immense variability in size, shape, and location. The model’s ability to accurately predict hemodynamics across all these variations, including rare or highly complex morphologies, needs thorough evaluation
Source: Nature
