One Disease, Two Worlds: An Explainable AI (XAI) Implementation for Interpretation of Risk Factors of Heart Failure Outcomes

Abstract

Machine learning can play an essential role in building a decision-making support system for use in clinical settings that helps quicken diagnosis and that can predict the risk of disease, in this case for heart failure. Unfortunately, a lack of trust from clinicians arises when the decision-making process of the models is not well explained, especially when the model applies to different demographics but is able to make a similar disease diagnosis. This study uses the SHAP method to explore the use of explainable artificial intelligence for model building through transfer learning with two different resources. We built a LightGBM predictive model that employs a novel cross-domain approach to improve a poor, low resource dataset by leveraging a rich, high resource dataset. This study revealed that this crossdomain approach enhanced the predictive performance of the model, with an absolute improvement of 12.86% in F1-score. Furthermore, the proposed approach still allowed the model to be interpreted with the XAI approach: 80% of the top 20 features of both cohorts that appeared in the feature importance analysis were overlap.

Ähnliche Arbeiten