Introduction – The discretized criteria for atrial electrogram (AEG) classi-fication as proposed by commercial systems (fractionated/not-fractionated) to guide ablation has been shown insufficient for persistent atrial fibrillation (persAF) therapy. In this study, we used unsupervised classification to inves-tigate possible sub-groups of persAF AEGs. Methods – 3745 bipolar AEGs were collected from 13 persAF patients af-ter pulmonary vein isolation. Automated AEG classification (fractionat-ed/not-fractionated) was performed using the CARTO criteria (Biosense Webster). The CARTO attributes (ICL, ACI and SCI) were used to create a 3D space attribute. K-mean with five groups and Euclidean distance was implemented. The five groups were compared with each other (median [IQR]). Results – Group 1 (43%) represents not-fractionated AEGs with low ICL (1.0 [1.0-2.0]), high ACI and SCI (110 [108-110] ms; 110 [107-110] ms). Group 2 (15%) has shown highly fractionated AEGs with high ICL (20 [17-24]), low ACI and SCI (75.0 [72.5-78.3] ms; 50.0 [50.0-50.8] ms). Groups 3 (9%) and 4 (9%) have shown similar low ICL (3 [1-5]; 2 [1-4]), but Group 3 has shown AEGs with short activation intervals (ACI 65.0 [57.5-69.2] ms; SCI 53.3 [50.8-57.5] ms), as opposed to Group 4 (ACI 87.5 [82.5-91.7] ms; SCI 78.3 [71.7-85.0] ms). Group 5 (23%) suggests moderated fractionation (ICL 8 [5-11]), with high ACI (80.0 [76.7-83.3] ms) but low SCI (54.2 [51.7-58.3] ms). The three attributes were significantly different among the five groups (P<0.0001), except ICL between Groups 3 and 4 (P>0.999) and SCI between Groups 3 and 5 (P>0.999). Conclusions – The results suggest that discretized AEG classification is in-sufficient to describe persAF and should be revisited. The unsupervised clas-sification proposed by the k-mean has found five sub-groups of AEGs with distinct characteristics, each of which could represent different electrophysio-logical mechanisms. This could provide a more detailed characterization of the atrial substrate during persAF ablation.