Cardiac arrhythmias present significant epidemiological and public health problems due to their relatively high prevalence in the general population. Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia treated in clinical practice. Most of the databases used for automated AF detection have an annotation indicating the beginning of a rhythm change during an AF episode, but no annotation for when that episode ends and normal sinus rhythm is restored. In this paper, from two databases with different progressive types of AF, over segments with 100 and 200 consecutive RR intervals extracted after the rhythm change annotation, heart rate variability features were obtained to examine how capturing parts of the ECG signal when the AF episode ends and normal sinus rhythm is restored affects the classification results. Three different machine learning classifiers were used on the sets of initially extracted features, statistically relevant features, and both statistically and clinically relevant features to examine how the classification results would be affected when irrelevant features are removed. A selection of clinically relevant features was performed through a brief literature review and clinical guidelines, while statistically relevant features were identified using appropriate statistical tests. Experimental results proved that capturing parts of the ECG signal when the AF episode ends and normal sinus rhythm is restored affects classification performance. In contrast, removing statistically and clinically irrelevant features not only maintains or improves classification performance but also reduces computational requirements, which is crucial for machine learning applications.