[Paper Review] Combining Residual U-Net and Data Augmentation for Dense Temporal Segmentation of Spike Wave Discharges in Single-Channel EEG

Saurav Sengupta, Scott Kilianski+5 more|arXiv (Cornell University)|Jan 1, 2026

EEG and Brain-Computer Interfaces0 citations

TL;DR

The paper introduces AugUNet1D, a residual 1D U-Net with data augmentation for dense temporal segmentation of spike-wave discharges in single-channel EEG, and shows it outperforms Twin Peaks and other baselines on a large mouse EEG dataset.

ABSTRACT

Manual annotation of spike-wave discharges (SWDs), the electrographic hallmark of absence seizures, is labor-intensive for long-term electroencephalography (EEG) monitoring studies. While machine learning approaches show promise for automated detection, they often struggle with cross-subject generalization due to high inter-individual variability in seizure morphology and signal characteristics. In this study we compare the performance of 15 machine learning classifiers on our own manually annotated dataset of 961 hours of EEG recordings from C3H/HeJ mice, including 22,637 labeled SWDs and find that a 1D U-Net performs the best. We then improve its performance by employing residual connections and data augmentation strategies combining amplitude scaling, Gaussian noise injection, and signal inversion during training to enhance cross-subject generalization. We also compare our method, named AugUNet1D, to a recently published time- and frequency-based algorithmic approach called "Twin Peaks" and show that AugUNet1D performs better on our dataset. AugUNet1D, pretrained on our manually annotated data or untrained, is made public for other users.

Motivation & Objective

• Address labor-intensive manual annotation of spike-wave discharges (SWDs) in long-term EEG recordings; investigate cross-subject generalization across variable SWD morphologies.
• Evaluate a broad set of classifiers to identify robust SWD detectors in C3H/HeJ mice.
• Develop enhancements via residual connections and data augmentation to improve cross-subject performance.
• Provide pretrained and untrained versions of AugUNet1D for broader applicability across electrographic events.

Proposed method

• Adopt a residual 1D U-Net architecture for dense, time-point level SWD segmentation.
• Apply residual connections within encoding/decoding blocks to improve gradient flow and train deeper models.
• Implement data augmentation including amplitude scaling, Gaussian noise injection, and signal inversion during training to improve robustness and generalization.
• Resample all EEG signals to 100 Hz and segment data into 20-second epochs (2000 samples) for training.
• Train with Dice Loss using Adam, with a cosine annealing warmup scheduler and early stopping for robust generalization.

Experimental results

Research questions

• RQ1 Can a residual 1D U-Net with specific data augmentations achieve robust cross-subject SWD detection in single-channel EEG?
• RQ2 How does AugUNet1D compare to a recently published time- and frequency-based method (Twin Peaks) on a large SWD dataset?
• RQ3 What training strategies best improve cross-subject generalization for dense temporal segmentation of SWDs in mouse EEG?
• RQ4 Is an untrained version of AugUNet1D useful for detecting diverse electrographic events across systems?

Key findings

• AugUNet1D achieved superior performance over a broad set of 16 baselines and the Twin Peaks approach on 10 test mice in their dataset.
• Data augmentation and residual connections jointly enhanced cross-subject generalization across recordings from 10 test mice at 100 Hz.
• The dataset comprised 961.3 hours of EEG with 22,637 manually labeled SWDs, demonstrating the scale and variability of the task.
• Manual SWD labeling focused on events with at least 5 rhythmic spike-wave complexes and regular intervals, providing a clear ground truth for evaluation.