Recurrence-Based Analysis of Electroencephalography Signals for Seizure Detection Using β-Divergences

Recurrence plots (RPs) are a powerful tool for visualizing and quantifying the dynamical structure of time series, particularly in biomedical signal analysis. Traditionally, RPs rely on fixed metrics such as the Euclidean distance to define recurrence. However, for complex, nonstationary, and noise-prone signals like electroencephalography (EEG), such metrics may fail to fully capture underlying dynamics or offer robustness in realistic settings. In this work, we introduce β-divergence as a flexible, parameterized alternative for RP construction. By interpolating between well-known divergences, such as Itakura-Saito (β=0) and Kullback-Leibler (β=1), as well as the Euclidean distance (β=2), this metric allows for tailored sensitivity to signal variation, offering an adjustable lens through which to enhance seizure detection in EEG.

We evaluate this method on EEG recordings from the Bonn University dataset, contrasting seizure (class E) and healthy (class A) brain states. After standard preprocessing steps (segmentation, bandpass filtering (delta, theta, alpha, beta), and normalization), RPs were constructed using 41 β values ranging from 0 to 5. To ensure comparability across β settings, we introduce a robust thresholding strategy based on normalized distance distributions and median absolute deviation.

From the β-divergence-based RPs, we extract 13 recurrence quantification analysis (RQA) features per frequency band (52 features in total), capturing nonlinear temporal dynamics. These features serve as input to a support vector machine (SVM) classifier trained to distinguish between seizure and non-seizure segments. We assess performance using the F1-score across both clean and noisy data, the latter simulated by adding Gaussian noise at multiple levels.

Results show that the best seizure detection performance is achieved with β = 1 (Kullback-Leibler divergence), reaching an F1-score of 95.3%, outperforming the Euclidean baseline (β = 2, 93%). Under noise, β = 3.5 and β = 1 remain top performers, maintaining F1-scores above 92.5%. Lower β values (β < 1) consistently underperform, especially in noisy conditions.

These findings demonstrate that β-divergence-based RPs can improve EEG-based seizure detection and offer better resilience to noise. The method holds promise for enhancing RP-based classification in medical signal analysis and other domains involving complex time-series data.