NOT ALL EEG MOMENTS ARE EQUAL: POSITION-ADAPTIVE TIME SCHEDULING FOR EEG GENERATION A PREPRINT

Introduction

Electroencephalography (EEG) records the electrical activity of the brain with high temporal resolution [van Ede et al., 2018]. EEG plays a central role in brain computer interfaces and neuroscience research [Perslev et al., 2021; Yogarajan et al., 2023; Lawhern et al., 2018; Zheng et al., 2019b].

High quality EEG data serves as a critical data source for the brain computer interface field [Yang et al., 2025].

EEG data collection is constrained by data scarcity due to high acquisition costs and expert annotation difficulties [You et al., 2025].

Research Question

We constrain the flow in the spectral domain via a multi-resolution STFT objective. The final training objective combines the four terms.

The sharp jump at the extreme tail supports our central hypothesis.

Future work may explore instance-level difficulty estimation that adapts to the content of each sample rather than relying solely on dataset-level statistics, extend position adaptive scheduling to multimodal and cross-subject settings, and investigate its applicability to other physiological time series.

Methodology

Our method achieves the largest accuracy gain on all three datasets. Our method consistently achieves the best classification performance on all three datasets.

Study Design

This analysis examines whether dataset-level difficulty yields disproportionate benefits.

Results & Findings

Identical effective supervision for easy and difficult regions prevents full utilization of limited model capacity. We identify overlooked heterogeneity as the central bottleneck limiting EEG generation fidelity.

• Identical effective supervision for easy and difficult regions prevents full utilization of limited model capacity.
• We identify overlooked heterogeneity as the central bottleneck limiting EEG generation fidelity.
• Position-Adaptive Time Scheduling (PATS) modulates position specific time progress based on historically observed difficulty.
• A Factorized Spatio-Temporal Attention module decomposes global self attention to explicitly model inter channel dependencies.
• EEG-GAN and ESC-GAN synthesize or edit EEG signals for augmentation [Hartmann et al., 2018; Fahimi et al., 2021; Zhang et al., 2024].

Related Work

This section reviews previous research in generative modeling, particularly focusing on the evolution from GANs to diffusion models and their applications in data synthesis, including EEG signal generation.

Generative Modeling for Data Synthesis

The section outlines the advancements in generative modeling techniques, emphasizing the limitations of GANs and the emergence of more stable diffusion and flow matching models for data synthesis.

EEG Signal Generation

This section details the trajectory of EEG signal generation methods, from GAN-based approaches to diffusion models, and discusses the limitations of existing methods in capturing the variability of real EEG recordings.

Figures Explained

The paper’s visual material highlights the workflow and the main system components.

• Figure 1 :: Figure 1: Illustration of the overlooked heterogeneity in EEG generation. Quasi stationary background activity (green) coexists with medium (yellow) and high difficulty (red) transient segments at different channel-time positions. Such patterns tend to recur at similar positions across recordings under the same protocol, motivating a position-wise difficulty statistic aggregated over the dataset rather than per instance.
• Figure 2 :: Figure 2: Overview of the proposed framework. (a) Training pipeline: PATS derives a position specific time map from the EMA error map, which conditions the Factorized Spatio-temporal Transformer to predict the clean sample. (b) Sampling pipeline: starting from Gaussian noise, PATS and the network are applied iteratively to solve the flow ODE and generate the final EEG output.
• Figure 3 :: Figure 3: Downstream classification performance (Accuracy, Macro-F1, Macro-Recall) on (a) TUEV, (b) BCIC-IV-2a, and (c) SEED-IV: real data only (No Aug) vs. real data augmented by each method. Error bars denote std over three seeds.
• Figure 4 :: Figure 4: Difficulty-stratified relative error reduction on TUEV. Instances are ranked by baseline (JET) error and grouped into ten equal-count deciles from easiest (1) to hardest (10); each point shows the mean relative error reduction of our framework over the baseline.