Multi-Modal And Multi-Approach Decision Support In Unruptured Intracranial Aneurysm
Newaz, Iftehaz (2025)
2025
Master's Programme in Biomedical Sciences and Engineering
Lääketieteen ja terveysteknologian tiedekunta - Faculty of Medicine and Health Technology
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-12-15
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025121511641
https://urn.fi/URN:NBN:fi:tuni-2025121511641
Tiivistelmä
Unruptured intracranial aneurysms (UIAs) affect roughly 3% of middle-aged adults. Although rupture is rare, it can lead to subarachnoid hemorrhage with high mortality (50%) and serious morbidity. Existing scoring systems and AI models provide useful risk estimates but often fail to identify aneurysms that later rupture and typically lack interpretability.
In this study, we developed machine learning models to predict rupture status and aneurysm growth using clinical and morphological features, and we introduced a hard voting ensemble method to improve predictions of future rupture. Among the models tested, Random Forest achieved the highest performance in distinguishing ruptured from unruptured aneurysms (AUC 98.8% ± 0.004), while the voting ensemble method provided better predictions for aneurysms that later ruptured which was previously diagnosed as unruptured aneurysm, a novel approach not previously explored in the literature.
To interpret the models, we applied explainable AI techniques including Partial Dependency Plots, feature importance analysis, and decision-tree visualizations. Across models, aneurysm geometry (length, diameter, neck width) was the strongest predictor of rupture and growth instability, while clinical factors such as smoking and prior subarachnoid hemorrhage also contributed for the prediction. Growth instability was further influenced by aneurysm location and patient age.
These findings demonstrate that interpretable machine learning combined with ensemble methods can capture meaningful patterns in aneurysm behaviour and provide patient-specific insights. However, accurately predicting future rupture remains challenging, partly due to the limited size of the dataset. This work establishes a new framework for ensemble-based prediction and interpretable evaluation of high-risk aneurysms.
In this study, we developed machine learning models to predict rupture status and aneurysm growth using clinical and morphological features, and we introduced a hard voting ensemble method to improve predictions of future rupture. Among the models tested, Random Forest achieved the highest performance in distinguishing ruptured from unruptured aneurysms (AUC 98.8% ± 0.004), while the voting ensemble method provided better predictions for aneurysms that later ruptured which was previously diagnosed as unruptured aneurysm, a novel approach not previously explored in the literature.
To interpret the models, we applied explainable AI techniques including Partial Dependency Plots, feature importance analysis, and decision-tree visualizations. Across models, aneurysm geometry (length, diameter, neck width) was the strongest predictor of rupture and growth instability, while clinical factors such as smoking and prior subarachnoid hemorrhage also contributed for the prediction. Growth instability was further influenced by aneurysm location and patient age.
These findings demonstrate that interpretable machine learning combined with ensemble methods can capture meaningful patterns in aneurysm behaviour and provide patient-specific insights. However, accurately predicting future rupture remains challenging, partly due to the limited size of the dataset. This work establishes a new framework for ensemble-based prediction and interpretable evaluation of high-risk aneurysms.