Epileptic seizure prediction from electroencephalogram (EEG) signals remains a challenging problem due to the high variability of brain dynamics and the strong class imbalance between interictal and preictal periods. This study presents a comparative evaluation of two adaptive filtering approaches, Kalman Filter and Recursive Least Squares (RLS), for early seizure prediction using a lightweight, real-time–oriented pipeline. Experiments were conducted on the CHB-MIT EEG database, comprising 684 recordings from 22 pediatric patients and 137 annotated seizures. EEG signals were segmented into overlapping windows, from which entropy, energy, variance, and mean absolute value were extracted, normalized, and aggregated into a compact feature representation. Seizure anticipation was formulated as an anomaly detection problem based on prediction error dynamics, with statistically defined thresholds and sustained alarm criteria. Results demonstrate that the Kalman Filter substantially outperforms RLS in seizure prediction rate, achieving 35.77% of seizures correctly anticipated with an average lead time of 84.2 s, whereas RLS predicted only 3.65% of events. Despite similar overall accuracy, the Kalman-based approach yielded more clinically meaningful alerts, with most detections occurring more than 60 s before seizure onset. The findings highlight the suitability of Kalman filtering for capturing preictal transitions through persistent innovation patterns, while RLS tends to absorb preictal changes into its adaptive model. Owing to its low computational complexity and real-time feasibility, the proposed Kalman-based framework represents a promising step toward embedded and wearable seizure prediction systems.