Introduction: Atrial fibrillation (AF) is the most common sustained arrhythmia and is associated with poor clinical outcomes, including stroke, acute coronary events, and heart failure. Recent studies have shown that early rhythm treatment of new-onset atrial fibrillation improves the patient's prognosis. However, atrial fibrillation is often asymptomatic, and it is difficult to determine its duration accurately. Recently, AI-based ECG has been studied for various cardiovascular diseases. Hypothesis: We sought to develop and validate a predictive model of the ECG for the duration of atrial fibrillation . Methods: All patients aged 18 years older from the two hospitals with at least one ECG were included in the study. Only patients with sinus rhythm with ECG prior to atrial fibrillation with ECG were selected. An ECG within 1 year from AF was first documented in ECG was defined as new onset AF. After dividing our datasets into training (and test sets, we developed an end-to-end deep neural network to predict for the duration of AF. Performance evaluation was conducted using various metrics, including AUROC, AUPRC, sensitivity, specificity, F1 score. Results: The dataset consisted of 83,525 ECGs from 16,193 patients from two hospitals. The AUROC for discriminating old AF and new-onset AF is 0.8186 (0.8181 - 0.8190) on internal validation set and 0.7967 (0.7966 - 0.7969) on external validation set. Sensitivity, Specificity, and F1 score are 0.7126((0.7118-0.7134), 0.7697 (0.7693-0.7701) and 0.5751 (0.5745-0.5757) on internal validation set and 0.7309 (0.7307-0.7311), 0.7225 (0.7224-0.7227) and 0.6354 (0.6352-0.6356)on external validation set. Conclusions: Our deep learning model can be used to predict atrial fibrillation's duration. Additional studies are ongoing to understand the relative importance of ECG features.

Introduction: Heart failure (HF) is a global pandemic with an increasing prevalence. The burden of HF-related hospitalizations and costs are increasing. Artificial intelligence(AI) algorithms applied to electrocardiograms have shown promise in diagnosing HF, but they require a large amount of training data, which is scarce and costly. Transfer learning addresses this challenge by utilizing knowledge from previous tasks, resulting in superior performance with limited data compared to conventional AI algorithms Hypothesis: AI algorithms using transfer learning can be accurately diagnose heart failure with reduced ejection fraction(HFrEF) using electrocardiograms compared to conventional AI algorithms. Methods: All patients aged 18 years older from the two hospitals with at least one ECG were included in the study. Electrocardiogram, transthoracic Echocardiogram (TTE), and demographic data were collected. The AI algorithm consisted of two phases: conventional AI algorithms, pre-training with a large-scale ECG dataset and transfer learning for HFrEF detection. Performance evaluation was conducted using various metrics, including AUROC, sensitivity, specificity, accuracy, F1 score. Results: The dataset consisted of 687,911 ECGs from 259,943 patients. Testing the conventional model on the HFrEF resulted in a sensitivity of 63.8%, specificity of 96.2%, accuracy of 93.6%, area under the receiver operating curve (AUC) of 0.93 (95% CI, 0.92 - 0.94), and F1 score of 61.6%. Testing the transfer learning model on the HFrEF increased the sensitivity of 75.1%, accuracy of 93.7%, area under the receiver operating curve (AUC) of 0.952 (95% CI, 0.93 - 0.96), and F1 score of 65.6%. but lowered the specificity of 95.3% Conclusions: A New AI-enabled ECG algorithm for identifying patients with HFrEF can be more accurately diagnosed than conventional AL algorithms.