Atrial fibrillation (AF) is the most common cardiac arrhythmia and can lead to sudden cardiac death and various complications, imposing a significant social burden. Electrocardiography is the gold standard for diagnosis, but missed and misdiagnoses of atrial fibrillation frequently occur due to variations in physicians' experience. Therefore, it is necessary to develop artificial intelligence models capable of accurately identifying AF in real-world clinically complex electrocardiogram (ECG) data.

To establish a clinical ECG database with multiple positive categories, and based on deep learning technology, utilize static ECG signal data to train a convolutional neural network (CNN) model for AF detection, aiming to improve the accuracy of automated AF diagnosis.

A total of 10 000 patients who met the inclusion and exclusion criteria and visited the Hunan Provincial People's Hospital (The First Affiliated Hospital of Hunan Normal University) in 2023 were selected. Among them, 1 462 cases were diagnosed with AF, and 8 538 cases were non-AF. General patient information (ages and genders) and static ECG data were collected to establish a clinical ECG database with multiple positive categories. The overall dataset was divided into training set (n=8 000), validation set (n=1 000), and test set (n=1 000) at 8∶1∶1 ratio. An advanced convolutional neural network model, AF Networks (AFNet), was developed using the training set for AF detection. The model's performance was evaluated using the validation and test sets, with metrics including sensitivity, positive predictive value, F1 score, accuracy, and the area under the ROC curve (AUC). These metrics were used to assess the model's performance in ECG diagnostic tasks and to analyze its strengths and limitations.

No statistically significant differences were observed in the comparisons across the three datasets for gender (χ²=1.32, P=0.517), age (F=0.87, P=0.419), and ECG distribution (χ²=2.666, P=0.264). In the test set, the AFNet model demonstrated a sensitivity of 98.00%, specificity of 99.29%, positive predictive value of 96.08%, negative predictive value of 99.65%, accuracy of 99.10%, an F1-score of 0.97, and AUC of 0.99 for diagnosing atrial fibrillation. In the validation set, the AFNet model also achieved high performance, with sensitivity, positive predictive value, accuracy, and F1-score reaching 96.81%, 95.80%, 99.67%, and 0.96, respectively.

The AFNet neural network model demonstrated efficient feature extraction for AF detection in a clinical ECG database with multiple positive categories. This is of great clinical value for AF screening.