Abstract

Background 
Intracranial aneurysms affect approximately 8% of the population, with a high incidence of subarachnoid hemorrhage This study analyses clinical data to investigate the risk factors associated with intracranial aneurysm rupture. The research utilizes a comprehensive dataset from the University Clinical Center in Gdańsk (UCCG), encompassing records from 2,095 patients hospitalized for both unruptured and ruptured aneurysms since 2006. The collected dataset was carefully prepared using data filtering, cleaning, standardizing, imputing, and aggregating methods. We conducted a multivariate and multidimensional exploratory statistical data analysis to differentiate between patients with ruptured and unruptured aneurysms, identifying key risk factors for aneurysm rupture. Finally, several ML models like LightGBM, XGBoost, and TabNet were trained, optimized, and validated to obtain the best possible efficacy of estimating the aneurysm rupture likelihood. 
Results 
The average intracranial aneurysm rupture probability measured on the test dataset through trained ML models ranges from 57.2% to 92.2%. In turn, sensitivity, which in binary classification describes the percentage of actual positive cases (aneurysm rupture cases) correctly identified by the model, ranges in our study from 0.61 to 1.0, indicating promising potential for clinical application. Based on the explainable AI methods, we indicate several features that significantly influence our ML models' predictions, i.g. glucose, platelets, creatinine, sodium, and patient age.  
Conclusion 
To the best of our knowledge, this is the first attempt to estimate the likelihood of intracranial aneurysm rupture based on routine laboratory test results and ML models. This study shows possible solutions to enhance diagnostic capabilities based on basic laboratory and clinical data collected routinely during hospitalization. This approach improves intracranial aneurysm patient outcomes and reduces healthcare costs associated with aneurysm management.