Introduction: Previous studies have demonstrated the usefulness of binary logistic regression equations in classifying the rupture status of cerebral aneurysms. However, the prediction accuracies that have been obtained remain short of the threshold necessary to become clinically relevant. Conventional approaches to improve accuracy include increasing the sample size, model complexity and variable considerations. This study aims to explore the utility of an alternative method. We employ a simple machine learning classifier on sample aneurysm data to evaluate the effectiveness of such a technique in predicting aneurysm rupture status.
Methods: Based on sample data from the literature, we created two datasets. Each dataset consisted of generated time-averaged wall shear stress (TAWSS) values and aspect ratios (AR) with corresponding ruptured or unruptured statuses. A support vector machine was trained and then tested using each dataset respectively. All data analyses were performed in MATLAB.
Results: The means and standard deviations of the generated data reflected values obtained from the literature. The training phase identified 4 support vectors for the unruptured population and 5 support vectors for the unruptured population. These support vectors were then used to identify a decision boundary which accurately separated the aneurysm populations by rupture status. During the testing phase, this support vector machine was able to classify 95% of the test data correctly.
Conclusions: Our results underscore the potential of machine learning classifiers. Further study should involve retrospective, patient-specific data and should additionally aim to include a larger sample size and additional variable considerations.
Patient Care: We hope to develop and refine computational methods that shed insight into the biophysical parameters that may affect aneurysm rupture. It is our belief that computational analyses could be used to better understand hemodynamic phenomena in a non-invasive manner. This improved understanding could then in turn be used to tailor treatment strategies that are optimized for each patient.
Learning Objectives: By the conclusion of this session, participants should be able to: (1) understand the theoretical underpinnings of machine learning algorithms; (2) evaluate the effectiveness of such algorithms when applied to test data; (3) consider potential applications for such algorithms in subsequent data analyses.