Abstract
Atrial Fibrillation (Afib) is a medical condition that makes the atria of the heart fibrillate, making the upper chambers of the heart(atria) beat in an uncoordinated manner with respect to ventricles. The electrically induced movement in the heart is called Rythm, when these electrical signals are disrupted, arrhythmia is triggered. This hinders the heart’s capacity to pump blood to the rest of the body as efficiently as it should and causes the formation of blood clots, thus increasing the chances of stroke. Atrial Fibrillation can be triggered after Cardiac and non-Cardiac surgery, which is described as Post-Operative Atrial Fibrillation (POAF). The pathological behavior of the disease is still not clear, which makes it even more important to diagnose. In this study, we introduce a novel Transformer-empowered framework to predict the onset of POAF after cardiac surgeries. Our framework is built on self-attention and co-attention using Transformers and consists of three major components: (1) the categorical and numeric feature encoder, (2) the self-attention embedding extraction, and (3) the co-attention embedding fusion. Specifically, the feature encoder converts each feature, categorical or numeric, into a unique token embedding. Then a stack of self-attention layers transforms categorical and numeric features, respectively, into their contextual embeddings. Subsequently, a stack of the co-attention layers captures the complex categorical-numeric interactions to obtain the robust fused categorical-numeric representations. Extensive experiments show that our new framework outperforms the state-of-the-art methods in terms of F1 score and AUC-ROC value.