Interpretable metrics for evaluating fidelity, diversity, and privacy in synthetic EEG data

Abstract

The generation of synthetic electroencephalography (EEG) data offers a promising strategy to mitigate the limited availability of clinical neurophysiological recordings, particularly intracranial EEG (iEEG), where acquisition is invasive, complex, and restricted to small patient cohorts. Although generative models can produce realistic signals, progress in this field is constrained by the absence of consistent and physiologically grounded evaluation practices. Current assessment strategies often rely on isolated statistical, visual, or downstream-task metrics, which may not fully capture the temporal, spectral, nonlinear, and spatial properties that characterize EEG signals. Here, we introduce seege_, an open-source Python library for evaluating synthetic EEG/iEEG data. It organizes complementary metrics into three evaluation dimensions: fidelity, diversity, and privacy. Fidelity is assessed across time, frequency, time–frequency, complexity, and spatial domains; diversity is evaluated through coverage, geometric structure, and intrinsic variability; and privacy is examined using nearest-neighbour similarity and membership-inference risk. To improve interpretability, the framework uses normalization strategies based on real-data variability, domain-level composite scores, and visual summaries that help identify agreement and failure modes. We illustrate this approach using synthetic intracranial EEG generated with a GAN-based model and validated through real–real, real–synthetic, synthetic–synthetic comparisons, surrogate controls, controlled perturbation experiments and generative failure modes. The results show that the proposed workflow captures domain-specific similarities and deviations between real and synthetic signals, supporting transparent reporting and comparison of synthetic EEG quality.