Background and objectives: Atrial fibrillation (AF) is a common and extremely harmful arrhythmia disease that leads to stroke and death. Several kinds of AF exist as paroxysmal AF, persistent AF, and permanent or chronic AF. Based on electrocardiograms (ECG), wavelet transform (WT), and convolutional neural networks (CNNs), the present study proposes an algorithm to classify between healthy subjects and patients suffering from paroxysmal atrial fibrillation (PAF). The classification results will then help for prediction purposes. Methods: The MIT-BIH “PAF” database (5-min signals before and during AF) and the PTB-XL database (each record is 10 s in duration) were used. Our model was tested on three leads of PTB-XL dataset. For each signal, a finite number of intrinsic mode functions (IMFs) was obtained using empirical mode decomposition. In order to select the most informative IMFs, two criteria were proposed: “variance ratio” and “percentage of occurrence of IMFs”. The selected IMFs serve as inputs to the continuous wavelet transform. The latter is used to extract features that are then inputs to our CNN classification model. The dataset was divided into training, validation, and test sets in proportions of 64%, 16%, and 20%, respectively. Results: For the “PAF” database, we achieve a maximum accuracy of 98.2% for the “variance ratio” selection and 98.8% for the “percentage of occurrence of IMFs”, compared to 97.1% and 97.5% before the application of these approaches. In addition, the presented approaches reduce the number of IMFs by at least 50%. The tested accuracy reached 98.8% on the PTB-XL dataset. Conclusion: By combining the variance ratio and percentages of IMFs, the extracted features allow to classify ECG of healthy patients from that of patients suffering from AF. The accuracy obtained outperforms some recent studies.