Long-term electrocardiograms (ECGS) recorded with mobile patches can help to detect paroxysmal diseases like atrial fibrillation (AF) when combined with automated ECG analysis. However, mobile patches provide a reduced number of leads that differ in signal morphology. We therefore investigated how reduced-lead ECGs affect AF detection, using 2,478 publicly available 12-lead ECGs of 30 s each. The feature set comprised 186 interpretable features per lead, including heart rate variability, morphology features, and signal quality indices. Binary decision tree ensembles were trained to detect AF, normal sinus rhythm (N), and other anomalies (O) in 8 different lead configurations. We also prospectively evaluated the applicability of the 3-lead model to 1,601 mobile long-term ECGs from the TIMELY eHealth project. Although the discriminability of AF, N, and O decreased with the number of leads, we achieved a minimum F1 score of 0.907 for single-lead III, compared with the highest F1 score of 0.957 when using 12-leads. P wave and PQ segment morphology features were consistently the most relevant. In mobile long-term ECGs, we were able to correctly identify AF in 5 patients, 2 of whom had no previous diagnosis. Overall, we achieved reliable classifications across different lead configurations and were able to demonstrate the potential of our approach for mobile applications.