1095 Systematic Machine Learning Approach for Cerebral Aneurysm Feature Selection and Rupture Status Classification

View at publisher Summarise with VIRAT
Authors: Lior Kofman, Calvin Ludwig, Alexandra Lauric, Emal Lesha, Adel M. Malek
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

INTRODUCTION: Machine learning techniques have increasing utility in clinical decision-making. Thus, a need for systematic, user-independent approaches to feature selection and classification is being sought for intracranial aneurysm rupture status classification. We propose such an approach using 8 classifiers. Previous literature utilizes datasets with small sample sizes, restricted to particular aneurysmal locations, and manually balanced ruptured/unruptured populations. METHODS: 3-dimensional angiograms of patient cerebral aneurysms (n = 678; 229 ruptured and 449 unruptured aneurysms) enabled computation of 13 morphological aneurysm features. After normalization of numerical features and univariate feature selection, 7:3 training:testing splitting was performed 100 times independently on each stratus. R package Caret feature selection was performed on training sets, followed by classification (8 classifiers). Area under the curve (AUC) performance was evaluated on testing sets. RESULTS: Neural networks boasted the greatest performance (AUCmean = 0.814, SENSmean = 0.750, SPECmean = 0.758), followed by multiple adaptive regression splines (MARS; AUCmean = 0.807, SENSmean = 0.756, SPECmean = 0.755). Sensitivities of these two models were found significant compared to those of the other classifiers. In neural networks, non-sphericity index, orthogonal height, and undulation index were identified as the strongest predictors of rupture status. In MARS models, aspect ratio and undulation index were computed as the strongest predictors. CONCLUSIONS: We propose a systematic, user-independent approach to machine learning aneurysm rupture status classification. Neural networks and MARS models exhibited the greatest AUC and sensitivities. Shape morphological factors such as non-sphericity index, aspect ratio, and undulation index were consistently selected as best descriptors.

Fields of Study & Subjects

Intracranial Aneurysms: Treatment and Complications Artificial Intelligence in Healthcare Aortic aneurysm repair treatments

Publication details

Journal / source
Neurosurgery
Volume
72
Issue
Supplement_1
Pages
204–204
Publication date
26 March 2026
Language
EN