Development and validation of stability prediction models for intracranial aneurysms based on ensemble learning...

This study aims to develop and validate ensemble learning models based on pre-rupture or pre-growth images andmultiple features to predict stability of intracranial aneurysm (IA).

Retrospectively collected aneurysms diagnosed through computed tomography angiography (CTA), magnetic resonance angiography (MRA) or digital subtraction angiography (DSA) from 7 hospitals. Internal dataset consists of data from Beijing Tiantan Hospital, and independent external validation set consists of data from the other 6 hospitals. We collected clinical information from medical records. Morphological and radiomics features were extracted from medical images. Each patient had at least one imaging follow-up. Predictive features are based on pre-rupture or pre-growth images. Univariate analysis, multivariate analysis and recursive feature elimination (RFE) algorithm were performed to select key features. We utilized ensemble learning algorithm to train models. Radiomics model, conventional model (based on clinical and morphological features) and combined model (based on radiomics, clinical and morphological features) were established and evaluated.

646 patients with 840 aneurysms were included from Beijing Tiantan Hospital as internal dataset, 206 patients with 271 aneurysms were included from the other 6 hospitals as independent external validation set. 15 radiomics features, 6 morphological features, and 2 clinical features were selected to build models. The radiomics model, conventional model and combined model obtained an AUC of 0.85(0.77-0.94), 0.61(0.48-0.74) and 0.78(0.67-0.89) in external validation set, respectively.

Radiomics features can be used to predict the stability of unruptured intracranial aneurysm (UIA). The combination of radiomics features and conventional features did not show the power to improve the predictive capability.