Toward Grading Subarachnoid Hemorrhage Risk Prediction: A Machine Learning-based Aneurysm Rupture Score (Preprint)

Abstract

Making clinical decisions about the treatment of Intracranial Aneurysms (IA) is not straightforward: small IAs in certain arteries may rupture while larger at others do not. Since many IA’s may not rupture for many years the risk of side effects from treatment prior to rupture must be weighed against the risk of SAH. Existing statistical and traditional approaches neither provide an accurate prediction of aneurysmal rupture nor offer a quantitative comparison among a group of SAH risk factors.

This paper evaluates the shortcomings of the PHASES and UIATS scores on retrospective data. Additionally, to address the complex challenge of SAH prediction, we develop an interpretable machine learning and data mining-based framework for the individualized rupture risk prediction of saccular intracranial aneurysm (SIA), and the aneurysmal rupture criticality prediction index for a set of risk factors to perform its relative comparison with other cases.

The proposed AI framework predicts the 5-year and lifetime rupture risk of intracranial aneurysms by training machine learning models based on an aneurysm’s location and shape. Additionally, it uses an ensemble learning model that is trained on data of all potential IA locations. Next, we employ the longitudinal data of 30 patients to develop a linear regression-based model to predict an aneurysm’s growth score. We use the Apriori algorithm to identify risk factors which carry a strong association with aneurysmal rupture for each combination of location and associated rupture risk factors. We compare the results not only with PHASES and UIATS scores but also with the scores of a multidisciplinary team of neurosurgeons.

The PHASES and UIATS scores show sensitivities of 22%, and 35%, and specificities of 76% and 79%, respectively. For the proposed framework, location-specific models show precision and recall of 93% and 90% for the Middle Cerebral Artery, 83% and 80% for the Anterior Communicating Artery, and 80% and 80% for the Supraclinoid Internal Carotid Artery. The ensemble method shows both precision and recall of 80%. The validation of the models on unseen data shows that the proposed framework performs better than our control group of neurosurgeons. Data-driven knowledge produces comparisons among 61 risk factor combinations, 11 ranked minor, 8 moderate, 41 severe, and one of which is a critical factor.

The PHASES and UIATS scores may only provide weak assistance in the clinical decision-making process of aneurysm treatment, particularly when it comes to the prediction of tiny aneurysms (those less than 5mm). The validation of the proposed framework, using an independent cohort (n=148), shows that it is possible to predict individualized rupture risk of saccular intracranial aneurysms of different sizes, including tiny ones. Apart from determining rupture probability in specific cases, the proposed Aneurysmal Rupture Criticality Prediction (ARCP) score could be used in training residents, both in understanding and in the ability to explain how the group of risk factors contribute to both 5-year and lifetime rupture risk.