Augmentation of EEG and ECG Time Series for Deep Learning Applications: Integrating Changepoint Detection into the iAAFT Surrogates
Nina Moutonnet 1 , Gregory Scott 2 , Danilo P. Mandic 3 · 2025
Read Augmentation of EEG and ECG Time Series for Deep Learning Applications: Integrating Changepoint Detection into the iAAFT Surrogates by Nina Moutonnet 1 , Gregory Scott 2 , Danilo P. Mandic 3 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-03761.txt.
Introduction
Nina Moutonnet, Gregory Scott, and Danilo Mandic present a method to expand brain and heart signal data in a 2025 paper. Deep learning needs much data, yet physiological signals are noisy and scarce. Their shifting patterns over time also break standard augmentation tricks.
Their method pairs changepoint detection with a Fourier-based surrogate technique. That pairing keeps the time and frequency traits of the original signal while making new variants. It was tested by training a seizure detector on original and expanded EEG data sets. Better augmentation could lift medical AI where collecting more real patient data is slow or sensitive.
Catalog record
| CREATOR | Nina Moutonnet 1 , Gregory Scott 2 , Danilo P. Mandic 3 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 30 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-03761
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Augmentation of EEG and ECG Time Series for Deep Learning Applications: Integrating Changepoint Detection into the iAAFT Surrogates Thanks: Nina Moutonnet was supported by the UKRI Centre for Doctoral Training in Artificial Intelligence for Healthcare (EP/S023283/1). Nina Moutonnet 1 , Gregory Scott 2 , Danilo P. Mandic 3 Abstract The performance of deep learning methods critically depends on the quality and quantity of the available training data. This is especially the case for physiological time series, which are both noisy and scarce, which calls for data augmentation to artificially increase the size of datasets. Another issue is that the time-evolving statistical properties of nonstationary signals prevent the use of standard data augmentation techniques. To this end, we introduce a novel method for augmenting nonstationary time series. This is achieved by combining offline changepoint detection with the iterative amplitude-adjusted Fourier transform (iAAFT), which ensures that the time-frequency properties of the original signal are preserved during augmentation. The proposed method is validated through comparisons of the performance of i) a deep learning seizure detection algorithm on both the original and augmented versions of the CHB-MIT and Siena scalp electroencephalography (EEG) databases, and ii) a deep learning atrial fibrillation (AF) detection algorithm on the original and augmented versions of the Computing in Cardiology Challenge 2017 dataset. By virtue of the proposed method, for the CHB-MIT and Siena datasets respectively, accuracy rose by 4.4% and 1.9%, precision by 10% and 5.5%, recall by 3.6% and 0.9%, and F1 by 4.2% and 1.4%. For the AF classification task, accuracy rose by 0.3%, precision by 2.1%, recall by 0.8%, and F1 by 2.1%. I Introduction The success of deep learning in the analysis of time series has also highlighted that large amount of training data is required to avoid overfitting and ensure generalisability. However, obtaining high quality labelled data can be challenging in real-world applications. In particular, medical data is often laborious and/or dangerous to collect, poorly labelled, and associated with data privacy concerns. Data augmentation is an effective way to artificially increase the size of existing datasets by creating synthetic copies which retain the statistical properties of the original data [ 1 , 2 ] . Whilst this has been extensively studied in computer vision, data augmentation for time series, such as in electroencephalography (EEG) and electrocardiography (ECG), is still in its infancy. In addition, temporal data augmentation is challenging due to the typically nonlinear and nonstationary nature of the data, lack of an intuitive understanding of the data, and unknown signal generating mechanisms. The lack of high-quality surrogate data is therefore an obstacle towards automatic analysis of EEG and ECG recordings using machine learning (ML), as open-source labelled datasets are still quite limited in size [ 3 ] . Current data augmentation strategies range from simple time-domain techniques through to generative models such as time-GAN [ 4 ] . These include time series manipulation, such as flipping along the vertical and horizontal axes [ 1 ] , downsampling [ 5 ] , and Gaussian noise injection [ 6 ] . More advanced techniques include dynamic time warping [ 7 ] , the iterative amplitude-adjusted Fourier transform (iAAFT) [ 5 ] , injecting complex patterns to a signal such as spikes and trends [ 8 , 9 ] , and generative AI models such as time-GAN and variational auto-encoders [ 10 , 11 ] . However, the choice of a data augmentation method that preserves the relevant characteristics of the original time series is not straightforward. For example, EEG and ECG recordings usually have a close to zero mean and do not exhibit long-term trends, but rather spikes and waves. Consequently, most simple signal augmenting manipulation techniques are not appropriate for EEG and ECG time series, e.g. time reversing a signal is only applicable to time series that are symmetric about the horizontal axis, which is not the case with EEG and ECG data. Similarly, while the iAAFT surrogates have equivalent first and second-order statistics to the original data [ 12 ] , they operate only on stationary data and are unsuitable for data augmentation of nonstationary time series. Generative AI models, despite their elegance, are extremely difficult to tune, and lack robustness as nonstationary data do not have a fixed probability density function. More specifically, the heterogeneity and sparsity of data poses a significant challenge when using DL-based augmentation strategies, a characteristic inherent to both ECG and EEG recordings. These physiological signals exhibit notable inter-patient variability, and their pathological manifestations encompass a broad spectrum of abnormalities. In the context of EEG, seizures may manifest as either generalised disturbances, present across all the channels, or as focal abnormalities localised to specific cortical regions and thus only visible in a subset of channels [ 3 ] . Similarly, cardiac arrhythmias, a type of atrial fibrillation (AF), demonstrate considerable heterogeneity in their electrophysiological expression, characterised by irregular R-R intervals and irregular morphological alterations in the PQRST complex [ 13 ] . To this end, the aim of this work is to help mitigate these issues by incorporating changepoint detection into data augmentation strategies. For EEG, this is achieved by assuming that the signal is composed of piecewise quasi-stationary segments. The existing nonstationary surrogates require prior knowledge of the changepoints of dynamical regimes in data, which is rarely available in practice [ 14 , 8 ] . The proposed data augmentation method, on the other hand, employs a changepoint detection scheme to segment the time series and applies a modified version of the iAAFT algorithm to each segment individually, which makes it suitable to nonstationary data. Experimental results in the context of deep learning for epileptic EEG support the analysis. Unlike the EEG augmentation approach, signal integrity during ECG augmentation is preserved through peak detection and retention rather than temporal segmentation. The peaks remain invariant during iAAFT augmentation, facilitating convergence to a surrogate that conserves the fundamental ECG morphology — and thus potentially significant clinical information — whilst introducing controlled variability in non-peak regions. II Proposed algorithm The proposed algorithm addresses nonstationarity through a two-stage process: initial detection of either quasi-stationary segment boundaries in EEG or characteristic peaks in ECG, followed by the use of the modified iAAFT algorithm for surrogate data generation. II-A Changepoint detection for EEG EEG signals exhibit temporally evolving signal statistics, due to variations in brain activity and/or the presence of artefacts [ 15 , 16 ] . This makes feature analysis a very challenging task in the characterisation and classification of EEG data. To this end, inspired by the work of Brodsky et al. [ 17 ] , a non-parametric offline changepoint detection method was devised based on the evolution of a set of features in the EEG signal. The work in [ 17 ] shows that detecting changes in a signal can be simplified to detecting changes in the expected value of an “engineered sequence” derived from the original data stream, termed a diagnostic sequence (x^{f}_{n})_{n\geq 1} , where f indexes the chosen feature. The task then boils down to selecting a subset of features which are known to discriminate well between seizure and non-seizure EEG segments and are also able to vary during a seizure. Each of such features corresponds to