Heart Rate (HR) data is obtained from ECG signal segments of different durations and used in the detection of different cardiac arrhythmias, including atrial fibrillation (AF). The duration of the ECG segments is directly linked to the number of consecutive RR intervals captured within. To obtain relevant heart rate variability (HRV) features, the appropriate duration of the ECG segments is often analyzed and evaluated. This is important when the classification of AF is based solely on the HRV features. In this paper, we examine whether different ECG segment durations affect the performance of four shallow machine learning classifiers (SVM, DT, RF, and KNN) used for binary classification into AF and normal sinus rhythm. The segment duration is measured in a number of consecutive RR intervals: 10, 20, 50, 100, and 200. The study is performed separately for the signals from the MIT-BIH Atrial Fibrillation and MIT-BIH Long Term Atrial Fibrillation databases. Performance metrics showed narrow variation both across different segment lengths and classifiers, especially for the MIT-BIH LTAF, with the accuracy and precision ranging in the interval 98.49% – 99.86% while for the MIT-BIH AF, the accuracy and precision interval is 95.73% – 100%. These experimental results indicate that the classification of the AF could be based on the shorter segments without a decrease in the performance of the classifier, thus enabling more efficient and timely detection of the AF episodes.