Evaluating Feature Prioritization in a Reduced Feature Set Using Explainable AI for Intracranial Aneurysm Classification

The timely management of an unruptured aneurysm can prevent hemorrhage. Multiple anatomical, morphological, and hemodynamic features associated with an aneurysm can predict the risk of rupture and classify the nature of the aneurysm as either ruptured or unruptured. However, considering all features may create ambiguities in the model due to inter-dependencies, such as the correlation between aneurysm height and width to aneurysm area. In computer-aided medical diagnostics, focusing on relevant features that diagnose the disease is more crucial than using an abundance of features that may not contribute to the diagnosis and may cause unnecessary complexity. The use of independent or sparsely correlated features improves the accuracy of the classifier. The proposed research analyzes the interdependency of features by calculating the correlation between features. Based on the analyses, the high correlation threshold is determined through rigorous experimentation. From the highly correlated features, one feature is selected. Consequently, the reduced feature set is recommended. The reduced feature set performs better with simpler classifiers. Instead of choosing a classifier that struggles with a broad range of features, an appropriately optimized feature set can enhance performance and improve accuracy. The proposed approach is tested on a publicly available dataset of aneurysms, leading to an improvement in classification accuracy from 82.6% to 87%, due to the use of an appropriate feature set. Moreover, the best model was evaluated using two different explainable AI frameworks to assess the contribution of each feature in prediction. However, the conflicting feature prioritization in each explainable framework suggests that these models prioritize different features for the same trained model, leading to ambiguity. This inconsistency indicates that the explainable models lack generalization and fail to consistently analyze feature importance.