Aiming to address the limitations of existing photoplethysmography (PPG)-based algorithms for atrial fibrillation (AF) detection—particularly their failure to account for the influence of non-AF arrhythmias such as premature ventricular contractions (PVC) and their lack of specificity in the presence of complex electrocardiogram (ECG) rhythms, this study proposes a high-specificity PPG algorithm for AF detection based on dual-channel feature fusion. The algorithm extracts PPG waveform features using a lightweight convolutional neural network (CNN), analyzes RR interval sequences through a long short-term memory (LSTM) network and identifies AF through feature fusion and hierarchical decision-making. On an independent test set, the method achieved a sensitivity, specificity and accuracy of 99.6%, 94.8%, and 96.3% respectively. The results demonstrate that the proposed algorithm effectively reduces the risk of misdiagnosis while maintaining a high detection rate in complex ECG backgrounds. It exhibits strong performance in AF identification and offers a more reliable screening solution for wearable health monitoring devices utilizing PPG technology.