Atrial fibrillation (AF) is a type of arrhythmia that can cause cardiac complications such as stroke, and early detection is therefore important. This study proposes a method for detecting AF from the Holter electrocardiogram (ECG). The AF detection procedure has two stages: arrhythmia extraction based on the R-R interval variation and AF identification from the extracted arrhythmia using a one-dimensional convolutional neural network (1D CNN). Artifacts in the ECG are eliminated through preprocessing using a finite-impulse-response bandpass filter. In the first stage, R waves are detected through a multi-resolution analysis of the ECG, and arrhythmias are extracted by observing the standard deviation of the R-R intervals. In the second stage, AF is identified from the extracted arrhythmic events using a 1D CNN trained using segmented ECG waveforms. An ECG dataset of 100,000 segments obtained from the Holter ECG is prepared for training the CNN. Evaluation using 24-h ECG data from 10 untrained subjects verifies that the performance of the proposed detection method is better than that of the methods without arrhythmia extraction, with an accuracy of 93.1%. This result indicates the feasibility of the proposed method for detecting AF.