In spite of the progress in management of Atrial Fibrillation (AF), this arrhythmia is one of the major causes of stroke and heart failure. The progression of this pathology from a silent paroxysmal form (PAF) into a sustained AF can be prevented by predicting the onset of PAF episodes. Moreover, since AF is caused by heterogeneous mechanisms in different patients, as we demonstrate in this paper, a patient-specific approach offers a promising solution. In this work, we consider two ECG recordings, one close to PAF onset and one far away from any PAF episode. For each patient, we extract two 5-minute ECG segments approximately 20 minutes apart. Next, we train a linear Support Vector Machine (SVM) classifier using patient-specific sets of time- and amplitude-domain features. In particular, we consider the P-waves and the QRS complexes in short windows of 5 consecutive heart beats. Finally, we validate the method on the PAF Prediction Challenge (2001) PhysioNet database predicting the onset with an F1 score of 97.1%, sensitivity of 96.2% and specificity of 98.1%.