Explainable AI (xAI) can identify samples that are relevant for deep neural networks (DNNs) to detect cardiovascular diseases in ECGs. However, interpretability is limited as these explanations are not always related to common diagnostic criteria. xECGArch comprises two convolutional neural networks (CNNs) with different temporal focus. As shown previously, the long-term CNN (LT-CNN) emphasizes QRS complexes, whose relevance changes correlate with rhythm. To test whether QRS morphology is decisive, we applied transfer learning to make the LT-CNN focus on rhythm for atrial fibrillation (AF) detection, using 10 s single-lead ECGs from public databases. The LT-CNN was trained on 9,675 ECGs to detect R peaks and tested on 1,320 unseen ECGs, reaching a sample-accurate F1 score of 98.1%. The pre-trained model was then fine-tuned on AF detection using 8,868 ECGs, with the weights of none or the first 3 to 8 of 9 layers frozen, reaching F1 scores of 87.6% to 93.3% on 986 unseen ECGs, decreasing with increasing number of frozen layers. A systematic validation of explanations extracted with deep Taylor decomposition shows increasing (p < 0.001) model focus on R peaks for AF detection with more frozen layers, peaking at 7. Our results indicate that transfer learning can guide DNNs to use specific features, enhancing interpretability and moving toward trustworthy AI for clinical applications.