Atrial flutter is a supraventricular reentrant arrhythmia that requires invasive electrical mapping for its correct characterization. The differentiation between the main variants — common and perimitral, both in clockwise or counterclockwise direction — is essential for planning ablation on the macroreentrant circuit. In this work, a non-invasive method based on the morphology of the F-wave of the ECG is proposed. By means of dimensionality reduction on the atrial signals and machine learning classifiers under stratified nested cross-validation, the four AFL subtypes are classified in a cohort of 150 patients. Three lead configurations are evaluated: 12 leads, the 4-lead subset (II, III, aVF, V1) and V1 in isolation. For the four-lead subset, the model obtained AUROCs of 0.84, 0.84, 0.88 and 0.72 for C-CCW, C-CW, PM-CCW and PM-CW respectively, matching or exceeding the full ECG in most subtypes. Notably, V1 in isolation maintained substantial performance, opening the possibility of preliminary subtype classification from a single lead. These results indicate that F-wave morphology, captured through a reduced set of leads and lower-dimensional representations, contains enough discriminative information for AFL subtype classification. Thus, we demonstrate that the F-wave contains morphological information to guide non-invasive ablation planning, laying the groundwork for future diagnostic support systems in the context of cardiac arrhythmias.