Analysis of Artificial Intelligence Methods in Classifying Heart Attack Risk: Black-Box Models vs. Glass-Box Models

Abstract

Artificial Intelligence (AI) is becoming more and more involved in human life day by day. Healthcare is one of the areas where AI is widely used, such as in the diagnosis prediction, and/or classification of diseases. Techniques such as machine learning provide high-accuracy results, but many algorithms have black-box structures, where the reasoning behind the predictions is not known. Explainable AI emerges to address this by providing explanations for complex models. While interpretable ("glass-box") models are desirable, they may have lower accuracy than complex ("black-box") models. Finding the right balance is crucial, especially in critical areas such as healthcare. It is also important to provide individual explanations for the predictions. This study uses patient data to explore a model to predict heart attack risk. Therefore, we compare glass-box models (logistic regression, naive Bayes, decision tree, and explainable boosting) with black-box models (random forest, support vector machine, multi-layer perceptron, gradient boosting, and stochastic gradient boosting). The results show that explainable boosting achieves the highest accuracy. To delve into individual explanations on a patient basis, the explainable boosting algorithm is compared with the random forest algorithm, which gives the best results among the black-box models. Here, LIME and SHAP are used to provide interpretability of random forests. As a result, it is concluded that the random forest algorithm has differences in the importance weights of the variables compared to the explainable boosting algorithm. Both results provide valuable tools for healthcare stakeholders to choose the most appropriate model.