This study was to assess the feasibility of using non-standardized single-lead electrocardiogram (ECG) monitoring to automatically detect atrial fibrillation (AF) with special emphasis on the combination of deep learning based algorithm and modified patch-based ECG lead. Fifty-five consecutive patients were monitored for AF in around 24 hours by patch-based ECG devices along with a standard 12-lead Holter. Catering to potential positional variability of patch lead, four typical positions on the upper-left chest were proposed. For each patch lead, the performance of automated algorithms with four different convolutional neural networks (CNN) was evaluated for AF detection against blinded annotations of two clinicians. A total of 349,388 10-second segments of AF and 161,084 segments of sinus rhythm were detected successfully. Good agreement between patch-based single-lead and standard 12-lead recordings was obtained at the position MP1 that corresponds to modified lead II, and a promising performance of the automated algorithm with an R-R intervals based CNN model was achieved on this lead in terms of accuracy (93.1%), sensitivity (93.1%), and specificity (93.4%). The present results suggest that the optimized patch-based ECG lead along by deep learning based algorithms may offer the possibility of providing an accurate, easy, and inexpensive clinical tool for mass screening of AF.