Abstract WMP28: A Novel Machine Learning Model for Rupture Prediction of Saccular Intracranial Aneurysm

Abstract

Background: Deciding how to deal with unruptured intracranial aneurysms is a complex problem. The research done so far has focused on identifying a contribution of individual risk factor (hypertension, female gender, etc.). However, it is difficult to translate the feature importance of single risk factor to a treatment decision. Recently, PHASES score was presented to predict rupture criticality, however, recent studies show it can't be used in clinical settings. Methods: A retrospective data of 915 patient records from an in-house registry of saccular aneurysms from 1997 - 2017 was used. We proposed a novel ensemble learning and apriori association rule mining-based method to define aneurysmal rupture risk in terms of Rupture Criticality Index (RCI), which is a normalized score (from 1 to 10) of the probability of aneurysmal rupture for set of associated features. In order to have interpretable and clinically useful model, we used Apriori algorithm to find what features carry strong association among them in the case of ruptured aneurysms for each combination of location and size. We used voting classifier as an ensemble technique and utilize weighted average probabilities (Soft Voting) to predict the outcome. Results: Proposed ensemble method shows precision, recall and f-measure of 75% and accuracy of 76%. The normalized values of RCI were categorized as minor, mild, moderate, severe, and critical by using Jenks natural breaks method. Evaluation results shows 26 ‘critical’, 53 ‘severe’ and 49 ‘moderate’ risk factors combinations. Examples of ‘critical’ combinations (which need immediate treatment) include: patients who are African American, current smoker, having anterior communicating artery (ACoA) Tiny-sized aneurysm (8.3-14.5 mm) with RCI of 8.57; patients who are African American, Male, having ACoA medium-sized aneurysm (8.3-14.5 mm) with RCI of 9.46; African American, Paraclinoid, Tiny-sized (<4.5mm) and current smoker with RCI 8.21, etc. Conclusion: The proposed machine learning model is capable of accurately predicting rupture risk over lifetime. Validation on retrospective data shows that results are reliable. Currently we are extending this work to make it as question-answer system using Google-home/Alexa voice assistants.