Driver drowsiness significantly contributes to road accidents worldwide, and timely prediction of driver reaction time is crucial for developing effective advanced driver assistance systems. In this paper, we present an EEG-based prediction framework that investigates the impact of different pre-stimulus time windows, frequency band combinations, and channel groups on driver reaction time estimation using data from a 90-minute sustained attention driving task. Our systematic evaluation using a publicly available dataset of 24 drivers reveals that a 2-second pre-stimulus window yields the lowest prediction error. Notably, our proposed 1D Convolutional Neural Network (CNN) approach reduces the Mean Absolute Error (MAE) by nearly 30\% (from 0.51 sec to 0.36 sec for the alpha band) compared to classical machine learning models. Moreover, while individual frequency bands (e.g., alpha and theta) outperform combined band approaches, most spatial channel groups deliver similar performance to the full 32-channel configuration-with the exception of frontal channels. These improvements underscore the potential for real-world applications in reducing road accidents by enabling timely interventions based on predictive analytics.