Explainable artificial intelligence in the prediction of high-risk asymptomatic carotid plaques based on ultrasonic image features.

Background: The addition of ultrasonic plaque texture features to clinical features in patients with asymptomatic internal carotid artery stenosis (ACS) improved the ability of a Support Vector Machine (SVM) model to identify plaques that are likely to produce stroke. However, SVM like many Artificial Intelligence (AI) black-box models lack transparency, limiting their adoption in critical settings. Explainable AI (XAI) techniques offer potential solutions by making model decision more interpretable. This study investigates whether incorporating XAI techniques can improve interpretability without significantly compromising predictive accuracy in stroke risk assessment.

Methods: We developed an Argumentation-based Explainable Machine Learning (ArgEML) methodology and framework for explainable machine learning predictions via argumentation. We used this framework to learn explainable argumentation theories from a real-life dataset of patients with asymptomatic carotid stenosis. We assessed the performance of these theories using standard machine learning (ML) metrics, while interpretability was evaluated through model transparency and quality of explanations.

Results: Results indicate that the ArgEML models maintain high predictive accuracy while significantly improving the interpretability of the predictions. Moreover, undecided predictions are addressed as dilemmas which still offer valuable information through the explanations of the different prediction capabilities.

Conclusions: Our findings suggest that ArgEML enhances the interpretability of stroke prediction from real life medical data without sacrificing predictive performance. Moreover, explanations offer valuable insights into misclassified cases and cases where a definite prediction cannot be derived. This transparency can help refine a model, guiding clinical decisions and improving AI adoption in healthcare.