Objective: Automatic detection of P-waves in electrocardiograms (ECGs) is difficult because of their small amplitudes and interference by ventricular activity and noise. This study aimed to improve P-wave detection using the periodic nature of sinoatrial (SA) node excitation. Methods: This study developed a phase regression deep neural network (PRDNN) that estimated the continuous phase of SA node-derived P-waves by assigning 0 rad at each P-wave onset. This architecture included bidirectional long short-term memory (BLSTM) layers, feature pyramid networks (FPN), and linear layers. The model was trained using 170 ECGs from 17 patients with second-degree atrioventricular (AV) blocks annotated by experts. Results: When tested on 22 ECGs (240 P-waves) from four AV block patients, the PRDNN achieved 98.8% recall, 99.2% precision, and F-measure of 0.990. The segmentation-based deep neural network (DNN) model used for comparison achieved 88.8% recall, 100% precision, and F-measure of 0.940. The PRDNN significantly improved sensitivity with a minimal loss in precision. Conclusion: By leveraging phase information, the proposed model improves the P-wave detection performance in automatic ECG analysis, potentially enhancing diagnostic support for clinicians.