Morphological parameters contributing to aneurysm rupture: Identifying the rogue from the mugshot using different statistical methods

INTRODUCTION: Current scores that discriminate rupture status/rupture-associated morphology of aneurysm rupture rely on size being >7 mm, ignoring important morphological parameters. This study applies advanced statistical methods on historical patient data to identify morphological features associated with aneurysms rupture. METHODS: Data from 429 aneurysms along with demographics and morphological data, collected by a blinded research assistant, were utilised to explore associations between morphology and demographic features and aneurysm ruptures. Associations between morphology and rupture status were assessed with generalised estimating equations clustered by patient. The relative importance of morphological and historical/demographic factors in predicting the risk of aneurysm rupture were investigated applying a selected machine learning techniques (random forest, neural network and support vector machine). RESULTS: Demographic characteristics were not statistically associated with risk of aneurysm rupture, but morphological characteristics (parent vessel diameter and aneurysm dome diameter) had significant associations with rupture. Accounting for only morphological parameters in predicting ruptured aneurysm, support vector machine was superior in detecting ruptured aneurysm compared to random forest (RF) and neural network. Considering morphological data together with patients' historical and demographic features, RF had the best performance with a substantial increase in the risk of rupture accuracy. In the latter model, age, aneurysm dome diameter, irregularity of shape, years of smoking and bottleneck factor were the most important variables in aneurysm rupture prediction. CONCLUSION: An aneurysm likely to rupture can be identified with a reasonable degree of certainty from a single scan using morphological characteristics alone although the accuracy of prediction increases significantly when other factors are included in the model. Larger studies to investigate this further are required.