Predicting postoperative atrial fibrillation: An explainable deep learning approach

Postoperative atrial fibrillation (POAF) is the most common complication after cardiac surgery.POAF causes a reduction in ventricular filling and atrial pooling, increases the risk of stroke threefold, and leads to other complications such as hypotension, pulmonary edema, acute kidney injury, heart failure, and encephalopathy [1][2] .In the present study, we developed TransAF, an accurate yet explainable transformer-powered deep learning system to predict the onset of POAF after cardiac surgery.By leveraging the transformer architecture [3][4] (the state-of-the-art deep learning technique for learning from heterogeneous, tabular data), TransAF represents a key step forward in bridging the gap between state-of-the-art deep learning research and the needs of the cardiac surgery community.The performance of TransAF and eight baseline predictive models was evaluated for comparison in terms of sensitivity, F1, and AUROC (the area under the receiver operating characteristic curve).Extensive experiments showed that TransAF outperformed all the state-of-the-art predictive methods with up to 46% and 15%improvements in terms of sensitivity and F1 scores, respectively.In addition to improved prediction performance, TransAF provides feature interaction analysis, feature attribution analysis, and feature importance analysis to support early intervention to optimize patient care by leveraging the latest gradient-based XAI (explainable artificial intelligence) algorithms.Specifically, to the best of our knowledge, TransAF constitutes the first attempt to explore the explainability of the POAF prediction results by revealing the complex, pairwise interactions among risk factors affecting POAF incidence through learning effective attention patterns using the transformer architecture [3][4] .We created a dataset from an institutional Society of Thoracic Surgeons (STS) database (records with missing values were excluded), which consisted of 2068 patient records with 149 features, of which 1587 were non-POAF patients and 481 were POAF patients.Patients with previously documented AF were excluded.Since our goal was to predict the onset of POAF after cardiac surgery, we excluded all postoperative features, which led to 82 features (77 categorical and 5 continuous).The definitions for all features are provided in Supplementary Tables 1 and 2. Demographic, clinical, and surgical characteristics of the dataset are summarized in Supplementary Tables 3 and 4.As shown in Fig. 1A, TransAF consists of three components: (1) a categorical and numerical feature encoder, (2) a self-attention embedding transformation, and (3) a co-attention embedding O r i g i n a l U n e d i t e d M a n u s c r i p t