Introduction: Ablation of persistent atrial fibrillation (persAF) targets using dominant frequency (DF), rotors, and complex fractionated atrial electrograms has been disappointing. A transfer learning technique applied to spectrograms may be a promising tool for predicting ablation outcomes. Methods: 3206 non-contact electrograms (EGMs) were collected for a time duration of 4 seconds before and after ablating 51 high DF locations of 10 patients with persAF. Two categories of data were labelled: 1490 EGMs (nodes) had positive ablation responses (AF termination or AF cycle length (AFCL) increased (≥10msec)), whereas 1716 EGMs had negative responses (AFCL increase (<10msec)) to catheter ablation. After the QRST subtraction process, EGMs were converted to spectrograms to visualize the variability of signals in the time-frequency domain. The residual network, equipped with a 50-layer pre-trained model, was utilized to extract features and train and test the transferred fully connected layers. The proposed model performance was evaluated by leaving EGMs of one patient out in a 10-fold cross-validation. Results: The 10-fold cross-validation accuracy, balanced accuracy, F1_score, AUC-ROC, sensitivity, specificity, and precision were 60.2%, 60.0%, 55.0%, 0.64, 51.5%, 67.8% and 58.2% respectively, based on the testing dataset. Conclusions: A transfer learning technique applied to features extracted from spectrograms might be useful to predict the responses of ablating electrograms and their effect on terminating AF and changes in CL.