Abstract
Post-operative atrial fibrillation (POAF) is a multi-faceted condition that contains a high degree of ambiguity in its risk assessment. Traditional prediction models like Random Forests (RF) and Support Vector Machines (SVM) were adopted to estimate POAF risk with varying degrees of success. However, these traditional methods often lack the ability to model complex–and often structural–interactions among multiple risk factors, which are highly important in clinical settings. This project aims to bridge the gap by developing ARNN (*) and ARNN (+), hybrid models of GCT and FT-Transformer to predict POAF probabilities and explain the interactions driving high-risk predictions. Extensive experiments using a real-world dataset showed that both ARNN (*) and ARNN (+) models outperformed traditional machine learning methods and state-of-the-art deep learning approaches including FT-Transformer and Tab-Transformer in terms of F1 Score.