A dual-stream multi-scale cross-attention network for pericardial effusion detection

Early detection of pericardial effusion (PE) is clinically important because delayed diagnosis may cause severe cardiac complications. We propose a Dual-Stream Multi-Scale Cross-Attention Network (DS-MSCA) for automatic PE detection using simultaneously acquired electrocardiogram (ECG) and phonocardiogram (PCG) signals. Multi-scale Inception-ResNet encoders capture waveform and rhythm variations, while bidirectional cross-attention models cardiac electromechanical coupling. On a clinical dataset of 149 patients, DS-MSCA achieved 93.26% accuracy, 95.28% sensitivity, and a 95.01% F1-score. Attention maps and Grad-CAM++ showed that the model focused on clinically relevant heart-sound abnormalities while integrating complementary ECG information, supporting non-invasive PE screening.