Seizure prediction is a critical challenge in healthcare, with significant implications for improving the quality of life of patients with epilepsy. Traditional methods often assume a fixed preictal period, typically 30 minutes, which may not adequately capture patient-specific and channel-specific variations. In this paper, we propose a novel approach to modeling preictal durations as probabilistic distributions and detecting change points in EEG time series using autoencoders. By leveraging the CHB-MIT Scalp EEG Database, we investigate preictal dynamics across multiple channels and evaluate the effectiveness of our method in identifying individualized preictal periods. Our approach segments EEG signals into fixed-size windows and trains autoencoders to detect deviations in reconstruction errors, which serve as indicators of change points. This framework captures complex temporal patterns, ensuring sensitivity to both abrupt and gradual transitions. Numerical experiments demonstrate that the assumed 30-minute preictal period exhibits significant variance, underscoring the necessity of adaptive modeling techniques. Analysis across 17 EEG channels highlights substantial variability in preictal durations, revealing the need for channel-specific detection strategies. Furthermore, we identify channels with higher rates of unpredicted events and propose weighting strategies to optimize model performance. The proposed method achieves stationary time series for preictal periods and aligns predicted preictal start times with observable transitions in EEG signals. Our findings emphasize the importance of personalized seizure prediction models that account for variability across patients and channels. By advancing the understanding of preictal dynamics and offering a robust change point detection framework, this study contributes to the development of reliable and scalable seizure prediction systems for diverse clinical applications.