Purpose: Non-invasively predicting the success of pharmacological cardioversion for patients with atrial fibrillation (AF) would be of great interest in clinical settings. In this study, an adaptive algorithm for tracking fundamental and harmonic components of atrial activity is proposed for extracting waveform features from surface ECG and discriminate patients with successful or failed cardioversion. Methods: 49 patients diagnosed with AF for whom pharmacological cardioversion was a success (15) or a failure (34) were studied. An adaptive tracking algorithm was applied to precordial ECG leads (V1-V6) for estimating the instantaneous frequency as well as extracting fundamental and first harmonic components with time-varying band-pass filters. Joint tracking on pairs of leads improved robustness. The phase difference between fundamental and harmonic signals was used as a measure of AF organization. Successful and failed cardioversions were classified with quadratic discriminant analysis based on the mean and variance of instantaneous AF frequency, and on the mean and variance of the phase difference slope. Results: The best selection of features for classifying successful and failed cardioversions achieved a correct rate of 81.6% with balanced sensitivity and specificity, using the mean instantaneous frequency and the mean and variance of phase difference slope estimated from leads V1 and V4. In this case, the negative predictive value was 90.3%, meaning that cardioversion failure could be predicted with high reliability. Conclusions: Adaptive tracking of AF harmonic components could potentially be used as a diagnostic tool for the assessment of future cardioversion efficacy. Indeed, an accurate prediction of future cardioversion failure could help tailoring treatments to appropriate options.