Atrial Fibrillation (AF) is the most common cardiac arrhythmia. Signal-processing approaches are widely used for the analysis of intracardiac electrograms (iEGMs), which are collected during catheter ablation from patients with AF. In order to identify possible targets for ablation therapy, dominant frequency (DF) is widely used and incorporated in electroanatomical mapping systems. Recently, a more robust measure, multiscale frequency (MSF), for iEGM data analysis was adopted and validated. However, before completing any iEGM analysis, a suitable bandpass (BP) filter must be applied to remove noise. Currently, no clear guidelines for BP filter characteristics exist. The lower bound of the BP filter is usually set to 3–5 Hz, while the upper bound (${\overline{BP}}^{th}$) of the BP filter varies from 15 Hz to 50 Hz according to many researchers. This large range of ${\overline{BP}}^{th}$ subsequently affects the efficiency of further analysis. In this paper, we aimed to develop a data-driven preprocessing framework for iEGM analysis, and validate it based on DF and MSF techniques. To achieve this goal, we optimized the ${\overline{BP}}^{th}$ using a data-driven approach (DBSCAN clustering) and demonstrated the effects of different ${\overline{BP}}^{th}$ on subsequent DF and MSF analysis of clinically recorded iEGMs from patients with AF. Our results demonstrated that our preprocessing framework with ${\overline{BP}}^{th}$ = 15 Hz has the best performance in terms of the highest Dunn index. We further demonstrated that the removal of noisy and contact-loss leads is necessary for performing correct data iEGMs data analysis.