Current classification of atrial fibrillation (AF) is crude, relying on AF duration and frequency, with little bearing on the severity of symptoms, risk of progression, and treatment success. Here, we use the Filter Diagonalization Method (FDM), a harmonic inversion technique first developed for characterizing molecular resonances in quantum physics, to extract f-wave features in electrocardiographic (ECG) traces for AF stratification. The FDM detects f-wave frequency, amplitude, and delay components at frame sizes of 0.15 seconds. Using features of the FDM outputs, we train statistical machine learning models to automatically learn discriminative dimensions in the feature vectors. The dataset comprised of ECG recordings from 22 patients before cryoablation; 2 had paroxysmal AF (PAF), 16 early persistent AF (Early-PeAF; <12 months duration), and 4 longstanding persistent AF (Longstanding-PeAF; >12 months duration). Of these, 2 PAF, 7 Early-PeAF and 1 Longstanding-PeAF received adenosine to enhance the R-R intervals prior to ablation. We randomly extracted 1500 f-waves from the pre-ablation without adenosine dataset, 500 each per AF category, to train Decision Tree (DT) and Random Forest (RF) classifiers. Ten-fold cross validation demonstrates that the RF performed best with accuracy 60.89%+/-0.31%. We also perform ten-fold cross-validation on 1474 feature vectors randomly extracted from the pre-ablation with adenosine dataset; for this, the DT performed best with classification accuracy 59.58%+/-0.04%. While the results are modest, they are better than chance (33%), showing that f-wave features can be used for AF stratification. Since current AF categories are crude, the results are expected to be moderate at best, and may well reflect disease progression and predict treatment response. The accuracies are similar for the two tests, slightly better without adenosine, showing that the FDM is able to successfully model short f-waves without the need to concatenate f-wave sequences nor the aid of adenosine to elongate R-R intervals.