Automated Detection of Patent Ductus Arteriosus in Pediatric Patients Using Doppler Ultrasonography Videos Based on a Transformer Model.

Patent ductus arteriosus (PDA) is a common congenital heart defect that requires timely and accurate detection to guide clinical management. Although deep learning has shown considerable promise in medical imaging, its application to echocardiographic video analysis remains challenging due to complex temporal dynamics and heterogeneous imaging conditions. TimeSformer, a Transformer-based architecture for temporal video modeling, is well suited for capturing long-range dependencies in echocardiographic sequences. In this study, we propose a novel two-stage artificial intelligence framework for automated PDA detection using Doppler echocardiography videos. In the first stage, parasternal short-axis (PSA) views are automatically identified and extracted from raw ultrasound videos. In the second stage, temporal features are analyzed to perform video-level diagnostic classification. To ensure robustness and generalizability, the proposed framework was developed and validated using a diverse multi-center dataset comprising examinations from four medical centers and four different ultrasound devices. The proposed method achieves high accuracy in view classification and effectively discriminates between PDA-positive and PDA-negative cases, yielding an area under the receiver operating characteristic curve (AUC) of 0.95. These results demonstrate the effectiveness of TimeSformer for echocardiographic sequence interpretation. Furthermore, the multi-center and multi-device validation highlights the adaptability of the framework, supporting its potential role as an AI-assisted diagnostic tool to enhance clinical workflows and patient outcomes in congenital heart disease.