Systematic machine learning approach for cerebral aneurysm feature selection and rupture status classification.

Area of Science:

• Neurosurgery
• Radiology
• Data Science

Background:

• Current intracranial aneurysm rupture risk assessment relies on manual measurements of morphological factors.
• Machine learning (ML) shows promise for improving aneurysm rupture risk stratification.
• Existing ML approaches often suffer from small sample sizes or anatomical restrictions.

Purpose of the Study:

• To develop and validate a computational tool for classifying intracranial aneurysm rupture status using ML.
• To assess the performance of various ML classifiers on a large, generalizable dataset of aneurysms.
• To identify key features contributing to aneurysm rupture risk.

Main Methods:

• Utilized 3D angiograms from 678 cerebral aneurysms (229 ruptured, 449 unruptured).
• Computed 21 aneurysm features, including morphological and locational data.
• Trained and tested 8 classifiers (7 ML models and 1 GLM) using 7:3 splits, evaluated by AUC, sensitivity, and specificity.

Main Results:

• Multiple adaptive regression splines (MARS) on morphological features achieved an AUC of 0.797.
• Neural networks (NNET) on morphological and locational features achieved the highest AUC of 0.825.
• Key predictors included aspect ratio, undulation index, and non-sphericity index.

Conclusions:

• A systematic, user-independent ML approach effectively classifies aneurysm rupture status.
• ML models, particularly NNET, outperform traditional generalized linear models (GLM).
• Morphological features like non-sphericity index, undulation index, and aspect ratio are crucial for risk assessment.