New ways to estimate and predict our health – improving the treatment of brain aneurysm with computational methods

CFD could help assess the risk of rupture

So far, CFD methods have seen little clinical use due to challenges in acquiring enough patient-specific data cost-effectively and the difficulty of validating the results. Additionally, the true connection between flow dynamics and aneurysm pathophysiology is yet to be revealed, which currently limits the potential for clinical use.

Progress has been made, however, as patient-specific geometries are already the industry standard. CFD can be used to predict phenomena such as loads on vessel walls, blood clotting, vortices and fluctuations in blood flow. Different flow-derived parameters linked with aneurysm formation, growth and rupture can also be utilized to determine the behavior of blood flow.

Challenges of blood flow simulation

Simulating blood flow is a particularly interesting topic due to its complexity. Blood is a non-Newtonian fluid, as it becomes runnier when it is stirred. Adding up to the challenge, the boundary conditions associated with blood flow are complex. They describe the conditions on the boundaries of the domain, such as vessel walls and flow inlets and outlets.

As is the case with all CFD models, correct boundary conditions are extremely important for producing a realistic solution. The boundary conditions are particularly complicated in this case, as the heart creates a pulsating flow and the vessel walls are distensible. Being able to accurately incorporate these features in the simulation model via boundary conditions is very difficult, as a great deal of complex medical measurements would have to be made to implement them.

The resulting simulation model is therefore quite complicated and incorporates some features rarely seen in other engineering CFD models. For example, in Elomatic’s pilot simulation, the outlet boundary conditions were implemented using the Windkessel model, meaning that for every flow outlet there is a small virtual electronic circuit. Circuit theory is then utilized to compute the transient boundary condition for each instant of a heartbeat.

Structural solver also required

Solving the velocity field over the course of a heartbeat is only one part of the whole solution, which includes the deformation of the vessel walls as well. This must be accomplished by coupling the CFD solver with a structural solver. This way, the CFD solver provides the flow field, while the structural solver computes the resulting deformation, creating a fluid-structure interaction.

Solving the vessel wall deformation accurately is challenging, as each person’s personal tissue properties would have to be measured or predicted. Factors such as age and plaque buildup affect vessel wall properties, adding up to the challenge.

Insights into events in the vascular system

As an outcome of the pilot simulations performed at Elomatic, we gained important insights into implementing the dynamic inlet and outlet boundary conditions and utilizing magnetic resonance imaging to produce patient-specific geometries using real patients. Different flow-derived parameters inside an aneurysm were analyzed in order to study the flow details.

The simulations also gave us insight into what could happen in the vascular system in terms of stresses and vortices when the fluid is considered non-Newtonian. Future research will focus on the implementation of fluid-structure interaction and more thorough patient-specific boundary conditions to provide a more realistic model. Eventually, the model needs to be validated to see how accurately it represents reality.

Mining patient data

Our next task was to train a machine learning model to predict the ICU outcome: death or discharge to hospice. VARHA provided us with a dataset containing thousands of features from brain aneurysm patients that was used to train the model.

In modern hospitals, a single patient visit generates a significant amount of data. While this data can contain valuable insights, the size and quality of the dataset can make it difficult to mine effectively. Fortunately, artificial intelligence-based solutions offer a powerful means of learning from large volumes of data.

Difficulty of producing a generalizable machine learning model

With thousands of features and only hundreds of patients, the risk of overfitting is high. This means that the resulting model can become closely tailored to this specific dataset and may not perform well with new data.

Furthermore, a vast majority (98%) of the features were missing, which is a common issue with this type of data. As an example, only half of the patients may have had their hemoglobin levels measured, while the other half may not, and therefore their values are missing. The missingness can itself be meaningful since it can reflect the physicians’ decisions. The time-series nature of the data, among other things, adds to the complexity of the dataset, making it crucial to carefully select the appropriate processing methods.

Excellent short-term predictions of ICU outcomes

As the goal of the research was to discover new predictors of ICU death for patients suffering from aneurysmal subarachnoid hemorrhage, we employed a combination of simple and complex feature selection techniques. This reduced the number of features from thousands to just a handful. These features were used to train multiple machine learning models, the results of which were then were compared.

The models demonstrated excellent performance in predicting ICU outcomes that occur within a week, although their performance decreased for longer-term outcomes. Overall, the model performance was good, but what was even more interesting was that the remaining features contained three predictors of death that had not been considered in prior machine learning-based research work.

Increasingly reliable early warning signs

Computational methods undoubtedly have their place in tomorrow’s healthcare: they can uncover new insights about diseases and enable physicians to peek into the future. In the case of ruptured and unruptured brain aneurysms, these methods can be used to assess the risk of rupture, to predict the likelihood of death in the event of a rupture, and to predict how well the patient will recover.

We have already continued the work with VARHA to repeat the data mining experiment with a larger patient population to validate the results. The newly discovered predictors can be combined with previously known predictors to improve the model’s performance. This will give clinicians more reliable early warning signs of poor patient outcomes.

These examples are just the tip of the iceberg of all the possible applications. Ultimately, the use of computational methods can lead to the development of more effective treatment options at every stage of the patient journey. In addition, the predictions generated by these methods can have an immediate impact on patients’ lives.

References

1. Vlak M.H., Algra A., Brandenburg R. & Rinkel G.J. (2011) Prevalence of unruptured intracranial aneurysms, with emphasis on sex, age, comorbidity, country, and time period: a systematic review and meta-analysis. The Lancet Neurology 10, pp. 626–636.
2. Geraghty J.R. & Testai F.D. (2017) Delayed cerebral ischemia after subarachnoid hemorrhage: beyond vasospasm and towards a multifactorial pathophysiology. Current atherosclerosis reports 19, pp. 1–12.
3. Van Gijn J., Kerr R.S. & Rinkel G.J. (2007) Subarachnoid haemorrhage. The Lancet 369, pp. 306–318.
4. Seibert B., Tummala R., Chow R., Faridar A., Mousavi S. & Divani A. (2011) Intracranial aneurysms: Review of current treatment options and outcomes. Frontiers in Neurology 2. URL: https://www.frontiersin.org/article/10.3389/fneur.2011.00045.
5. Tawk R.G., Hasan T.F., D’Souza C.E., Peel J.B. & Freeman W.D. (2021) Diagnosis and treatment of unruptured intracranial aneurysms and aneurysmal subarachnoid hemorrhage. Mayo Clinic Proceedings 96, pp. 1970–2000. URL: https://www.sciencedirect.com/science/ article/pii/S0025619621000410.